/*

 * Copyright © 2018-2021, 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 <errno.h>

#include <limits.h>

#include <string.h>

#include <stdio.h>

#include <inttypes.h>

#include "dav1d/data.h"

#include "common/frame.h"

#include "common/intops.h"

#include "src/ctx.h"

#include "src/decode.h"

#include "src/dequant_tables.h"

#include "src/env.h"

#include "src/filmgrain.h"

#include "src/log.h"

#include "src/qm.h"

#include "src/recon.h"

#include "src/ref.h"

#include "src/tables.h"

#include "src/thread_task.h"

#include "src/warpmv.h"

static void init_quant_tables(const Dav1dSequenceHeader *const seq_hdr,

                              const Dav1dFrameHeader *const frame_hdr,

                              const int qidx, uint16_t (*dq)[3][2])

{

    for (int i = 0; i < (frame_hdr->segmentation.enabled ? 8 : 1); i++) {

        const int yac = frame_hdr->segmentation.enabled ?

            iclip_u8(qidx + frame_hdr->segmentation.seg_data.d[i].delta_q) : qidx;

        const int ydc = iclip_u8(yac + frame_hdr->quant.ydc_delta);

        const int uac = iclip_u8(yac + frame_hdr->quant.uac_delta);

        const int udc = iclip_u8(yac + frame_hdr->quant.udc_delta);

        const int vac = iclip_u8(yac + frame_hdr->quant.vac_delta);

        const int vdc = iclip_u8(yac + frame_hdr->quant.vdc_delta);

        dq[i][0][0] = dav1d_dq_tbl[seq_hdr->hbd][ydc][0];

        dq[i][0][1] = dav1d_dq_tbl[seq_hdr->hbd][yac][1];

        dq[i][1][0] = dav1d_dq_tbl[seq_hdr->hbd][udc][0];

        dq[i][1][1] = dav1d_dq_tbl[seq_hdr->hbd][uac][1];

        dq[i][2][0] = dav1d_dq_tbl[seq_hdr->hbd][vdc][0];

        dq[i][2][1] = dav1d_dq_tbl[seq_hdr->hbd][vac][1];

    }

}

static int read_mv_component_diff(MsacContext *const msac,

                                  CdfMvComponent *const mv_comp,

                                  const int mv_prec)

{

    const int sign = dav1d_msac_decode_bool_adapt(msac, mv_comp->sign);

    const int cl = dav1d_msac_decode_symbol_adapt16(msac, mv_comp->classes, 10);

    int up, fp = 3, hp = 1;

    if (!cl) {

        up = dav1d_msac_decode_bool_adapt(msac, mv_comp->class0);

        if (mv_prec >= 0) {  // !force_integer_mv

            fp = dav1d_msac_decode_symbol_adapt4(msac, mv_comp->class0_fp[up], 3);

            if (mv_prec > 0) // allow_high_precision_mv

                hp = dav1d_msac_decode_bool_adapt(msac, mv_comp->class0_hp);

        }

    } else {

        up = 1 << cl;

        for (int n = 0; n < cl; n++)

            up |= dav1d_msac_decode_bool_adapt(msac, mv_comp->classN[n]) << n;

        if (mv_prec >= 0) {  // !force_integer_mv

            fp = dav1d_msac_decode_symbol_adapt4(msac, mv_comp->classN_fp, 3);

            if (mv_prec > 0) // allow_high_precision_mv

                hp = dav1d_msac_decode_bool_adapt(msac, mv_comp->classN_hp);

        }

    }

    const int diff = ((up << 3) | (fp << 1) | hp) + 1;

    return sign ? -diff : diff;

}

static void read_mv_residual(Dav1dTileState *const ts, mv *const ref_mv,

                             const int mv_prec)

{

    MsacContext *const msac = &ts->msac;

    const enum MVJoint mv_joint =

        dav1d_msac_decode_symbol_adapt4(msac, ts->cdf.mv.joint, N_MV_JOINTS - 1);

    if (mv_joint & MV_JOINT_V)

        ref_mv->y += read_mv_component_diff(msac, &ts->cdf.mv.comp[0], mv_prec);

    if (mv_joint & MV_JOINT_H)

        ref_mv->x += read_mv_component_diff(msac, &ts->cdf.mv.comp[1], mv_prec);

}

static void read_tx_tree(Dav1dTaskContext *const t,

                         const enum RectTxfmSize from,

                         const int depth, uint16_t *const masks,

                         const int x_off, const int y_off)

{

    const Dav1dFrameContext *const f = t->f;

    const int bx4 = t->bx & 31, by4 = t->by & 31;

    const TxfmInfo *const t_dim = &dav1d_txfm_dimensions[from];

    const int txw = t_dim->lw, txh = t_dim->lh;

    int is_split;

    if (depth < 2 && from > (int) TX_4X4) {

        const int cat = 2 * (TX_64X64 - t_dim->max) - depth;

        const int a = t->a->tx[bx4] < txw;

        const int l = t->l.tx[by4] < txh;

        is_split = dav1d_msac_decode_bool_adapt(&t->ts->msac,

                       t->ts->cdf.m.txpart[cat][a + l]);

        if (is_split)

            masks[depth] |= 1 << (y_off * 4 + x_off);

    } else {

        is_split = 0;

    }

    if (is_split && t_dim->max > TX_8X8) {

        const enum RectTxfmSize sub = t_dim->sub;

        const TxfmInfo *const sub_t_dim = &dav1d_txfm_dimensions[sub];

        const int txsw = sub_t_dim->w, txsh = sub_t_dim->h;

        read_tx_tree(t, sub, depth + 1, masks, x_off * 2 + 0, y_off * 2 + 0);

        t->bx += txsw;

        if (txw >= txh && t->bx < f->bw)

            read_tx_tree(t, sub, depth + 1, masks, x_off * 2 + 1, y_off * 2 + 0);

        t->bx -= txsw;

        t->by += txsh;

        if (txh >= txw && t->by < f->bh) {

            read_tx_tree(t, sub, depth + 1, masks, x_off * 2 + 0, y_off * 2 + 1);

            t->bx += txsw;

            if (txw >= txh && t->bx < f->bw)

                read_tx_tree(t, sub, depth + 1, masks,

                             x_off * 2 + 1, y_off * 2 + 1);

            t->bx -= txsw;

        }

        t->by -= txsh;

    } else {

        dav1d_memset_pow2[t_dim->lw](&t->a->tx[bx4], is_split ? TX_4X4 : txw);

        dav1d_memset_pow2[t_dim->lh](&t->l.tx[by4], is_split ? TX_4X4 : txh);

    }

}

static int neg_deinterleave(int diff, int ref, int max) {

    if (!ref) return diff;

    if (ref >= (max - 1)) return max - diff - 1;

    if (2 * ref < max) {

        if (diff <= 2 * ref) {

            if (diff & 1)

                return ref + ((diff + 1) >> 1);

            else

                return ref - (diff >> 1);

        }

        return diff;

    } else {

        if (diff <= 2 * (max - ref - 1)) {

            if (diff & 1)

                return ref + ((diff + 1) >> 1);

            else

                return ref - (diff >> 1);

        }

        return max - (diff + 1);

    }

}

static void find_matching_ref(const Dav1dTaskContext *const t,

                              const enum EdgeFlags intra_edge_flags,

                              const int bw4, const int bh4,

                              const int w4, const int h4,

                              const int have_left, const int have_top,

                              const int ref, uint64_t masks[2])

{

    /*const*/ refmvs_block *const *r = &t->rt.r[(t->by & 31) + 5];

    int count = 0;

    int have_topleft = have_top && have_left;

    int have_topright = imax(bw4, bh4) < 32 &&

                        have_top && t->bx + bw4 < t->ts->tiling.col_end &&

                        (intra_edge_flags & EDGE_I444_TOP_HAS_RIGHT);

#define bs(rp) dav1d_block_dimensions[(rp)->bs]

#define matches(rp) ((rp)->ref.ref[0] == ref + 1 && (rp)->ref.ref[1] == -1)

    if (have_top) {

        const refmvs_block *r2 = &r[-1][t->bx];

        if (matches(r2)) {

            masks[0] |= 1;

            count = 1;

        }

        int aw4 = bs(r2)[0];

        if (aw4 >= bw4) {

            const int off = t->bx & (aw4 - 1);

            if (off) have_topleft = 0;

            if (aw4 - off > bw4) have_topright = 0;

        } else {

            unsigned mask = 1 << aw4;

            for (int x = aw4; x < w4; x += aw4) {

                r2 += aw4;

                if (matches(r2)) {

                    masks[0] |= mask;

                    if (++count >= 8) return;

                }

                aw4 = bs(r2)[0];

                mask <<= aw4;

            }

        }

    }

    if (have_left) {

        /*const*/ refmvs_block *const *r2 = r;

        if (matches(&r2[0][t->bx - 1])) {

            masks[1] |= 1;

            if (++count >= 8) return;

        }

        int lh4 = bs(&r2[0][t->bx - 1])[1];

        if (lh4 >= bh4) {

            if (t->by & (lh4 - 1)) have_topleft = 0;

        } else {

            unsigned mask = 1 << lh4;

            for (int y = lh4; y < h4; y += lh4) {

                r2 += lh4;

                if (matches(&r2[0][t->bx - 1])) {

                    masks[1] |= mask;

                    if (++count >= 8) return;

                }

                lh4 = bs(&r2[0][t->bx - 1])[1];

                mask <<= lh4;

            }

        }

    }

    if (have_topleft && matches(&r[-1][t->bx - 1])) {

        masks[1] |= 1ULL << 32;

        if (++count >= 8) return;

    }

    if (have_topright && matches(&r[-1][t->bx + bw4])) {

        masks[0] |= 1ULL << 32;

    }

#undef matches

}

static void derive_warpmv(const Dav1dTaskContext *const t,

                          const int bw4, const int bh4,

                          const uint64_t masks[2], const union mv mv,

                          Dav1dWarpedMotionParams *const wmp)

{

    int pts[8][2 /* in, out */][2 /* x, y */], np = 0;

    /*const*/ refmvs_block *const *r = &t->rt.r[(t->by & 31) + 5];

#define add_sample(dx, dy, sx, sy, rp) do { \

    pts[np][0][0] = 16 * (2 * dx + sx * bs(rp)[0]) - 8; \

    pts[np][0][1] = 16 * (2 * dy + sy * bs(rp)[1]) - 8; \

    pts[np][1][0] = pts[np][0][0] + (rp)->mv.mv[0].x; \

    pts[np][1][1] = pts[np][0][1] + (rp)->mv.mv[0].y; \

    np++; \

} while (0)

    // use masks[] to find the projectable motion vectors in the edges

    if ((unsigned) masks[0] == 1 && !(masks[1] >> 32)) {

        const int off = t->bx & (bs(&r[-1][t->bx])[0] - 1);

        add_sample(-off, 0, 1, -1, &r[-1][t->bx]);

    } else for (unsigned off = 0, xmask = (uint32_t) masks[0]; np < 8 && xmask;) { // top

        const int tz = ctz(xmask);

        off += tz;

        xmask >>= tz;

        add_sample(off, 0, 1, -1, &r[-1][t->bx + off]);

        xmask &= ~1;

    }

    if (np < 8 && masks[1] == 1) {

        const int off = t->by & (bs(&r[0][t->bx - 1])[1] - 1);

        add_sample(0, -off, -1, 1, &r[-off][t->bx - 1]);

    } else for (unsigned off = 0, ymask = (uint32_t) masks[1]; np < 8 && ymask;) { // left

        const int tz = ctz(ymask);

        off += tz;

        ymask >>= tz;

        add_sample(0, off, -1, 1, &r[off][t->bx - 1]);

        ymask &= ~1;

    }

    if (np < 8 && masks[1] >> 32) // top/left

        add_sample(0, 0, -1, -1, &r[-1][t->bx - 1]);

    if (np < 8 && masks[0] >> 32) // top/right

        add_sample(bw4, 0, 1, -1, &r[-1][t->bx + bw4]);

    assert(np > 0 && np <= 8);

#undef bs

    // select according to motion vector difference against a threshold

    int mvd[8], ret = 0;

    const int thresh = 4 * iclip(imax(bw4, bh4), 4, 28);

    for (int i = 0; i < np; i++) {

        mvd[i] = abs(pts[i][1][0] - pts[i][0][0] - mv.x) +

                 abs(pts[i][1][1] - pts[i][0][1] - mv.y);

        if (mvd[i] > thresh)

            mvd[i] = -1;

        else

            ret++;

    }

    if (!ret) {

        ret = 1;

    } else for (int i = 0, j = np - 1, k = 0; k < np - ret; k++, i++, j--) {

        while (mvd[i] != -1) i++;

        while (mvd[j] == -1) j--;

        assert(i != j);

        if (i > j) break;

        // replace the discarded samples;

        mvd[i] = mvd[j];

        memcpy(pts[i], pts[j], sizeof(*pts));

    }

    if (!dav1d_find_affine_int(pts, ret, bw4, bh4, mv, wmp, t->bx, t->by) &&

        !dav1d_get_shear_params(wmp))

    {

        wmp->type = DAV1D_WM_TYPE_AFFINE;

    } else

        wmp->type = DAV1D_WM_TYPE_IDENTITY;

}

static inline int findoddzero(const uint8_t *buf, int len) {

    for (int n = 0; n < len; n++)

        if (!buf[n * 2]) return 1;

    return 0;

}

// meant to be SIMD'able, so that theoretical complexity of this function

// times block size goes from w4*h4 to w4+h4-1

// a and b are previous two lines containing (a) top/left entries or (b)

// top/left entries, with a[0] being either the first top or first left entry,

// depending on top_offset being 1 or 0, and b being the first top/left entry

// for whichever has one. left_offset indicates whether the (len-1)th entry

// has a left neighbour.

// output is order[] and ctx for each member of this diagonal.

static void order_palette(const uint8_t *pal_idx, const ptrdiff_t stride,

                          const int i, const int first, const int last,

                          uint8_t (*const order)[8], uint8_t *const ctx)

{

    int have_top = i > first;

    assert(pal_idx);

    pal_idx += first + (i - first) * stride;

    for (int j = first, n = 0; j >= last; have_top = 1, j--, n++, pal_idx += stride - 1) {

        const int have_left = j > 0;

        assert(have_left || have_top);

#define add(v_in) do { \

        const int v = v_in; \

        assert((unsigned)v < 8U); \

        order[n][o_idx++] = v; \

        mask |= 1 << v; \

    } while (0)

        unsigned mask = 0;

        int o_idx = 0;

        if (!have_left) {

            ctx[n] = 0;

            add(pal_idx[-stride]);

        } else if (!have_top) {

            ctx[n] = 0;

            add(pal_idx[-1]);

        } else {

            const int l = pal_idx[-1], t = pal_idx[-stride], tl = pal_idx[-(stride + 1)];

            const int same_t_l = t == l;

            const int same_t_tl = t == tl;

            const int same_l_tl = l == tl;

            const int same_all = same_t_l & same_t_tl & same_l_tl;

            if (same_all) {

                ctx[n] = 4;

                add(t);

            } else if (same_t_l) {

                ctx[n] = 3;

                add(t);

                add(tl);

            } else if (same_t_tl | same_l_tl) {

                ctx[n] = 2;

                add(tl);

                add(same_t_tl ? l : t);

            } else {

                ctx[n] = 1;

                add(imin(t, l));

                add(imax(t, l));

                add(tl);

            }

        }

        for (unsigned m = 1, bit = 0; m < 0x100; m <<= 1, bit++)

            if (!(mask & m))

                order[n][o_idx++] = bit;

        assert(o_idx == 8);

#undef add

    }

}

static void read_pal_indices(Dav1dTaskContext *const t,

                             uint8_t *const pal_idx,

                             const int pal_sz, const int pl,

                             const int w4, const int h4,

                             const int bw4, const int bh4)

{

    Dav1dTileState *const ts = t->ts;

    const ptrdiff_t stride = bw4 * 4;

    assert(pal_idx);

    uint8_t *const pal_tmp = t->scratch.pal_idx_uv;

    pal_tmp[0] = dav1d_msac_decode_uniform(&ts->msac, pal_sz);

    uint16_t (*const color_map_cdf)[8] =

        ts->cdf.m.color_map[pl][pal_sz - 2];

    uint8_t (*const order)[8] = t->scratch.pal_order;

    uint8_t *const ctx = t->scratch.pal_ctx;

    for (int i = 1; i < 4 * (w4 + h4) - 1; i++) {

        // top/left-to-bottom/right diagonals ("wave-front")

        const int first = imin(i, w4 * 4 - 1);

        const int last = imax(0, i - h4 * 4 + 1);

        order_palette(pal_tmp, stride, i, first, last, order, ctx);

        for (int j = first, m = 0; j >= last; j--, m++) {

            const int color_idx = dav1d_msac_decode_symbol_adapt8(&ts->msac,

                                      color_map_cdf[ctx[m]], pal_sz - 1);

            pal_tmp[(i - j) * stride + j] = order[m][color_idx];

        }

    }

    t->c->pal_dsp.pal_idx_finish(pal_idx, pal_tmp, bw4 * 4, bh4 * 4,

                                 w4 * 4, h4 * 4);

}

static void read_vartx_tree(Dav1dTaskContext *const t,

                            Av1Block *const b, const enum BlockSize bs,

                            const int bx4, const int by4)

{

    const Dav1dFrameContext *const f = t->f;

    const uint8_t *const b_dim = dav1d_block_dimensions[bs];

    const int bw4 = b_dim[0], bh4 = b_dim[1];

    // var-tx tree coding

    uint16_t tx_split[2] = { 0 };

    b->max_ytx = dav1d_max_txfm_size_for_bs[bs][0];

    if (!b->skip && (f->frame_hdr->segmentation.lossless[b->seg_id] ||

                     b->max_ytx == TX_4X4))

    {

        b->max_ytx = b->uvtx = TX_4X4;

        if (f->frame_hdr->txfm_mode == DAV1D_TX_SWITCHABLE) {

            dav1d_memset_pow2[b_dim[2]](&t->a->tx[bx4], TX_4X4);

            dav1d_memset_pow2[b_dim[3]](&t->l.tx[by4], TX_4X4);

        }

    } else if (f->frame_hdr->txfm_mode != DAV1D_TX_SWITCHABLE || b->skip) {

        if (f->frame_hdr->txfm_mode == DAV1D_TX_SWITCHABLE) {

            dav1d_memset_pow2[b_dim[2]](&t->a->tx[bx4], b_dim[2 + 0]);

            dav1d_memset_pow2[b_dim[3]](&t->l.tx[by4], b_dim[2 + 1]);

        }

        b->uvtx = dav1d_max_txfm_size_for_bs[bs][f->cur.p.layout];

    } else {

        assert(bw4 <= 16 || bh4 <= 16 || b->max_ytx == TX_64X64);

        int y, x, y_off, x_off;

        const TxfmInfo *const ytx = &dav1d_txfm_dimensions[b->max_ytx];

        for (y = 0, y_off = 0; y < bh4; y += ytx->h, y_off++) {

            for (x = 0, x_off = 0; x < bw4; x += ytx->w, x_off++) {

                read_tx_tree(t, b->max_ytx, 0, tx_split, x_off, y_off);

                // contexts are updated inside read_tx_tree()

                t->bx += ytx->w;

            }

            t->bx -= x;

            t->by += ytx->h;

        }

        t->by -= y;

        if (DEBUG_BLOCK_INFO)

            printf("Post-vartxtree[%x/%x]: r=%d\n",

                   tx_split[0], tx_split[1], t->ts->msac.rng);

        b->uvtx = dav1d_max_txfm_size_for_bs[bs][f->cur.p.layout];

    }

    assert(!(tx_split[0] & ~0x33));

    b->tx_split0 = (uint8_t)tx_split[0];

    b->tx_split1 = tx_split[1];

}

static inline unsigned get_prev_frame_segid(const Dav1dFrameContext *const f,

                                            const int by, const int bx,

                                            const int w4, int h4,

                                            const uint8_t *ref_seg_map,

                                            const ptrdiff_t stride)

{

    assert(f->frame_hdr->primary_ref_frame != DAV1D_PRIMARY_REF_NONE);

    unsigned seg_id = 8;

    ref_seg_map += by * stride + bx;

    do {

        for (int x = 0; x < w4; x++)

            seg_id = imin(seg_id, ref_seg_map[x]);

        ref_seg_map += stride;

    } while (--h4 > 0 && seg_id);

    assert(seg_id < 8);

    return seg_id;

}

static inline void splat_oneref_mv(const Dav1dContext *const c,

                                   Dav1dTaskContext *const t,

                                   const enum BlockSize bs,

                                   const Av1Block *const b,

                                   const int bw4, const int bh4)

{

    const enum InterPredMode mode = b->inter_mode;

    const refmvs_block ALIGN(tmpl, 16) = (refmvs_block) {

        .ref.ref = { b->ref[0] + 1, b->interintra_type ? 0 : -1 },

        .mv.mv[0] = b->mv[0],

        .bs = bs,

        .mf = (mode == GLOBALMV && imin(bw4, bh4) >= 2) | ((mode == NEWMV) * 2),

    };

    c->refmvs_dsp.splat_mv(&t->rt.r[(t->by & 31) + 5], &tmpl, t->bx, bw4, bh4);

}

static inline void splat_intrabc_mv(const Dav1dContext *const c,

                                    Dav1dTaskContext *const t,

                                    const enum BlockSize bs,

                                    const Av1Block *const b,

                                    const int bw4, const int bh4)

{

    const refmvs_block ALIGN(tmpl, 16) = (refmvs_block) {

        .ref.ref = { 0, -1 },

        .mv.mv[0] = b->mv[0],

        .bs = bs,

        .mf = 0,

    };

    c->refmvs_dsp.splat_mv(&t->rt.r[(t->by & 31) + 5], &tmpl, t->bx, bw4, bh4);

}

static inline void splat_tworef_mv(const Dav1dContext *const c,

                                   Dav1dTaskContext *const t,

                                   const enum BlockSize bs,

                                   const Av1Block *const b,

                                   const int bw4, const int bh4)

{

    assert(bw4 >= 2 && bh4 >= 2);

    const enum CompInterPredMode mode = b->inter_mode;

    const refmvs_block ALIGN(tmpl, 16) = (refmvs_block) {

        .ref.ref = { b->ref[0] + 1, b->ref[1] + 1 },

        .mv.mv = { b->mv[0], b->mv[1] },

        .bs = bs,

        .mf = (mode == GLOBALMV_GLOBALMV) | !!((1 << mode) & (0xbc)) * 2,

    };

    c->refmvs_dsp.splat_mv(&t->rt.r[(t->by & 31) + 5], &tmpl, t->bx, bw4, bh4);

}

static inline void splat_intraref(const Dav1dContext *const c,

                                  Dav1dTaskContext *const t,

                                  const enum BlockSize bs,

                                  const int bw4, const int bh4)

{

    const refmvs_block ALIGN(tmpl, 16) = (refmvs_block) {

        .ref.ref = { 0, -1 },

        .mv.mv[0].n = INVALID_MV,

        .bs = bs,

        .mf = 0,

    };

    c->refmvs_dsp.splat_mv(&t->rt.r[(t->by & 31) + 5], &tmpl, t->bx, bw4, bh4);

}

static void mc_lowest_px(int *const dst, const int by4, const int bh4,

                         const int mvy, const int ss_ver,

                         const struct ScalableMotionParams *const smp)

{

    const int v_mul = 4 >> ss_ver;

    if (!smp->scale) {

        const int my = mvy >> (3 + ss_ver), dy = mvy & (15 >> !ss_ver);

        *dst = imax(*dst, (by4 + bh4) * v_mul + my + 4 * !!dy);

    } else {

        int y = (by4 * v_mul << 4) + mvy * (1 << !ss_ver);

        const int64_t tmp = (int64_t)(y) * smp->scale + (smp->scale - 0x4000) * 8;

        y = apply_sign64((int)((llabs(tmp) + 128) >> 8), tmp) + 32;

        const int bottom = ((y + (bh4 * v_mul - 1) * smp->step) >> 10) + 1 + 4;

        *dst = imax(*dst, bottom);

    }

}

static ALWAYS_INLINE void affine_lowest_px(Dav1dTaskContext *const t, int *const dst,

                                           const uint8_t *const b_dim,

                                           const Dav1dWarpedMotionParams *const wmp,

                                           const int ss_ver, const int ss_hor)

{

    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;

    const int y = b_dim[1] * v_mul - 8; // lowest y

    const int src_y = t->by * 4 + ((y + 4) << ss_ver);

    const int64_t mat5_y = (int64_t) mat[5] * src_y + mat[1];

    // check left- and right-most blocks

    for (int x = 0; x < b_dim[0] * h_mul; x += imax(8, b_dim[0] * h_mul - 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 int64_t mvy = ((int64_t) mat[4] * src_x + mat5_y) >> ss_ver;

        const int dy = (int) (mvy >> 16) - 4;

        *dst = imax(*dst, dy + 4 + 8);

    }

}

static NOINLINE void affine_lowest_px_luma(Dav1dTaskContext *const t, int *const dst,

                                           const uint8_t *const b_dim,

                                           const Dav1dWarpedMotionParams *const wmp)

{

    affine_lowest_px(t, dst, b_dim, wmp, 0, 0);

}

static NOINLINE void affine_lowest_px_chroma(Dav1dTaskContext *const t, int *const dst,

                                             const uint8_t *const b_dim,

                                             const Dav1dWarpedMotionParams *const wmp)

{

    const Dav1dFrameContext *const f = t->f;

    assert(f->cur.p.layout != DAV1D_PIXEL_LAYOUT_I400);

    if (f->cur.p.layout == DAV1D_PIXEL_LAYOUT_I444)

        affine_lowest_px_luma(t, dst, b_dim, wmp);

    else

        affine_lowest_px(t, dst, b_dim, wmp, f->cur.p.layout & DAV1D_PIXEL_LAYOUT_I420, 1);

}

static void obmc_lowest_px(Dav1dTaskContext *const t,

                           int (*const dst)[2], const int is_chroma,

                           const uint8_t *const b_dim,

                           const int bx4, const int by4, const int w4, const int h4)

{

    assert(!(t->bx & 1) && !(t->by & 1));

    const Dav1dFrameContext *const f = t->f;

    /*const*/ refmvs_block **r = &t->rt.r[(t->by & 31) + 5];

    const int ss_ver = is_chroma && f->cur.p.layout == DAV1D_PIXEL_LAYOUT_I420;

    const int ss_hor = is_chroma && f->cur.p.layout != DAV1D_PIXEL_LAYOUT_I444;

    const int h_mul = 4 >> ss_hor, v_mul = 4 >> ss_ver;

    if (t->by > t->ts->tiling.row_start &&

        (!is_chroma || 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_block *const a_r = &r[-1][t->bx + x + 1];

            const uint8_t *const a_b_dim = dav1d_block_dimensions[a_r->bs];

            if (a_r->ref.ref[0] > 0) {

                const int oh4 = imin(b_dim[1], 16) >> 1;

                mc_lowest_px(&dst[a_r->ref.ref[0] - 1][is_chroma], t->by,

                             (oh4 * 3 + 3) >> 2, a_r->mv.mv[0].y, ss_ver,

                             &f->svc[a_r->ref.ref[0] - 1][1]);

                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_block *const l_r = &r[y + 1][t->bx - 1];

            const uint8_t *const l_b_dim = dav1d_block_dimensions[l_r->bs];

            if (l_r->ref.ref[0] > 0) {

                const int oh4 = iclip(l_b_dim[1], 2, b_dim[1]);

                mc_lowest_px(&dst[l_r->ref.ref[0] - 1][is_chroma],

                             t->by + y, oh4, l_r->mv.mv[0].y, ss_ver,

                             &f->svc[l_r->ref.ref[0] - 1][1]);

                i++;

            }

            y += imax(l_b_dim[1], 2);

        }

}

static int decode_b(Dav1dTaskContext *const t,

                    const enum BlockLevel bl,

                    const enum BlockSize bs,

                    const enum BlockPartition bp,

                    const enum EdgeFlags intra_edge_flags) {

    Dav1dTileState *const ts = t->ts;

    const Dav1dFrameContext *const f = t->f;

    Av1Block b_mem, *const b = t->frame_thread.pass ?

        &f->frame_thread.b[t->by * f->b4_stride + t->bx] : &b_mem;

    const uint8_t *const b_dim = dav1d_block_dimensions[bs];

    const int bx4 = t->bx & 31, by4 = t->by & 31;

    const int ss_ver = f->cur.p.layout == DAV1D_PIXEL_LAYOUT_I420;

    const int ss_hor = f->cur.p.layout != DAV1D_PIXEL_LAYOUT_I444;

    const int cbx4 = bx4 >> ss_hor, cby4 = by4 >> ss_ver;

    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 cbw4 = (bw4 + ss_hor) >> ss_hor, cbh4 = (bh4 + ss_ver) >> ss_ver;

    const int have_left = t->bx > ts->tiling.col_start;

    const int have_top = t->by > ts->tiling.row_start;

    const int has_chroma = f->cur.p.layout != DAV1D_PIXEL_LAYOUT_I400 &&

                           (bw4 > ss_hor || t->bx & 1) &&

                           (bh4 > ss_ver || t->by & 1);

    if (t->frame_thread.pass == 2) {

        if (b->intra) {

            f->bd_fn.recon_b_intra(t, bs, intra_edge_flags, b);

            const enum IntraPredMode y_mode_nofilt =

                b->y_mode == FILTER_PRED ? DC_PRED : b->y_mode;

#define set_ctx(rep_macro) \

            rep_macro(edge->mode, off, y_mode_nofilt); \

            rep_macro(edge->intra, off, 1)

            BlockContext *edge = t->a;

            for (int i = 0, off = bx4; i < 2; i++, off = by4, edge = &t->l) {

                case_set(b_dim[2 + i]);

            }

#undef set_ctx

            if (IS_INTER_OR_SWITCH(f->frame_hdr)) {

                refmvs_block *const r = &t->rt.r[(t->by & 31) + 5 + bh4 - 1][t->bx];

                for (int x = 0; x < bw4; x++) {

                    r[x].ref.ref[0] = 0;

                    r[x].bs = bs;

                }

                refmvs_block *const *rr = &t->rt.r[(t->by & 31) + 5];

                const int bx = t->bx + bw4 - 1;

                for (int y = 0; y < bh4 - 1; y++) {

                    refmvs_block *d = &rr[y][bx];

                    d->ref.ref[0] = 0;

                    d->bs = bs;

                }

            }

            if (has_chroma) {

                uint8_t uv_mode = b->uv_mode;

                dav1d_memset_pow2[ulog2(cbw4)](&t->a->uvmode[cbx4], uv_mode);

                dav1d_memset_pow2[ulog2(cbh4)](&t->l.uvmode[cby4], uv_mode);

            }

        } else {

            if (IS_INTER_OR_SWITCH(f->frame_hdr) /* not intrabc */ &&

                b->comp_type == COMP_INTER_NONE && b->motion_mode == MM_WARP)

            {

                if (b->matrix[0] == INT16_MIN) {

                    t->warpmv.type = DAV1D_WM_TYPE_IDENTITY;

                } else {

                    t->warpmv.type = DAV1D_WM_TYPE_AFFINE;

                    t->warpmv.matrix[2] = b->matrix[0] + 0x10000;

                    t->warpmv.matrix[3] = b->matrix[1];

                    t->warpmv.matrix[4] = b->matrix[2];

                    t->warpmv.matrix[5] = b->matrix[3] + 0x10000;

                    dav1d_set_affine_mv2d(bw4, bh4, b->mv2d, &t->warpmv,

                                          t->bx, t->by);

                    dav1d_get_shear_params(&t->warpmv);

#define signabs(v) v < 0 ? '-' : ' ', abs(v)

                    if (DEBUG_BLOCK_INFO)

                        printf("[ %c%x %c%x %c%x\n  %c%x %c%x %c%x ]\n"

                               "alpha=%c%x, beta=%c%x, gamma=%c%x, delta=%c%x, mv=y:%d,x:%d\n",

                               signabs(t->warpmv.matrix[0]),

                               signabs(t->warpmv.matrix[1]),

                               signabs(t->warpmv.matrix[2]),

                               signabs(t->warpmv.matrix[3]),

                               signabs(t->warpmv.matrix[4]),

                               signabs(t->warpmv.matrix[5]),

                               signabs(t->warpmv.u.p.alpha),

                               signabs(t->warpmv.u.p.beta),

                               signabs(t->warpmv.u.p.gamma),

                               signabs(t->warpmv.u.p.delta),

                               b->mv2d.y, b->mv2d.x);

#undef signabs

                }

            }

            if (f->bd_fn.recon_b_inter(t, bs, b)) return -1;

            const uint8_t *const filter = dav1d_filter_dir[b->filter2d];

            BlockContext *edge = t->a;

            for (int i = 0, off = bx4; i < 2; i++, off = by4, edge = &t->l) {

#define set_ctx(rep_macro) \

                rep_macro(edge->filter[0], off, filter[0]); \

                rep_macro(edge->filter[1], off, filter[1]); \

                rep_macro(edge->intra, off, 0)

                case_set(b_dim[2 + i]);

#undef set_ctx

            }

            if (IS_INTER_OR_SWITCH(f->frame_hdr)) {

                refmvs_block *const r = &t->rt.r[(t->by & 31) + 5 + bh4 - 1][t->bx];

                const int ref1 = b->ref[0] + 1;

                const union mv mv1 = b->mv[0];

                for (int x = 0; x < bw4; x++) {

                    r[x].ref.ref[0] = ref1;

                    r[x].mv.mv[0] = mv1;

                    r[x].bs = bs;

                }

                refmvs_block *const *rr = &t->rt.r[(t->by & 31) + 5];

                const int bx = t->bx + bw4 - 1;

                for (int y = 0; y < bh4 - 1; y++) {

                    refmvs_block* d = &rr[y][bx];

                    d->ref.ref[0] = ref1;

                    d->mv.mv[0] = mv1;

                    d->bs = bs;

                }

            }

            if (has_chroma) {

                dav1d_memset_pow2[ulog2(cbw4)](&t->a->uvmode[cbx4], DC_PRED);

                dav1d_memset_pow2[ulog2(cbh4)](&t->l.uvmode[cby4], DC_PRED);

            }

        }

        return 0;

    }

    const int cw4 = (w4 + ss_hor) >> ss_hor, ch4 = (h4 + ss_ver) >> ss_ver;

    b->bl = bl;

    b->bp = bp;

    b->bs = bs;

    const Dav1dSegmentationData *seg = NULL;

    // segment_id (if seg_feature for skip/ref/gmv is enabled)

    int seg_pred = 0;

    if (f->frame_hdr->segmentation.enabled) {

        if (!f->frame_hdr->segmentation.update_map) {

            if (f->prev_segmap) {

                unsigned seg_id = get_prev_frame_segid(f, t->by, t->bx, w4, h4,

                                                       f->prev_segmap,

                                                       f->b4_stride);

                if (seg_id >= 8) return -1;

                b->seg_id = seg_id;

            } else {

                b->seg_id = 0;

            }

            seg = &f->frame_hdr->segmentation.seg_data.d[b->seg_id];

        } else if (f->frame_hdr->segmentation.seg_data.preskip) {

            if (f->frame_hdr->segmentation.temporal &&

                (seg_pred = dav1d_msac_decode_bool_adapt(&ts->msac,

                                ts->cdf.m.seg_pred[t->a->seg_pred[bx4] +

                                t->l.seg_pred[by4]])))

            {

                // temporal predicted seg_id

                if (f->prev_segmap) {

                    unsigned seg_id = get_prev_frame_segid(f, t->by, t->bx,

                                                           w4, h4,

                                                           f->prev_segmap,

                                                           f->b4_stride);

                    if (seg_id >= 8) return -1;

                    b->seg_id = seg_id;

                } else {

                    b->seg_id = 0;

                }

            } else {

                int seg_ctx;

                const unsigned pred_seg_id =

                    get_cur_frame_segid(t->by, t->bx, have_top, have_left,

                                        &seg_ctx, f->cur_segmap, f->b4_stride);

                const unsigned diff = dav1d_msac_decode_symbol_adapt8(&ts->msac,

                                          ts->cdf.m.seg_id[seg_ctx],

                                          DAV1D_MAX_SEGMENTS - 1);

                const unsigned last_active_seg_id =

                    f->frame_hdr->segmentation.seg_data.last_active_segid;

                b->seg_id = neg_deinterleave(diff, pred_seg_id,

                                             last_active_seg_id + 1);

                if (b->seg_id > last_active_seg_id) b->seg_id = 0; // error?

                if (b->seg_id >= DAV1D_MAX_SEGMENTS) b->seg_id = 0; // error?

            }

            if (DEBUG_BLOCK_INFO)

                printf("Post-segid[preskip;%d]: r=%d\n",

                       b->seg_id, ts->msac.rng);

            seg = &f->frame_hdr->segmentation.seg_data.d[b->seg_id];

        }

    } else {

        b->seg_id = 0;

    }

    // skip_mode

    if ((!seg || (!seg->globalmv && seg->ref == -1 && !seg->skip)) &&

        f->frame_hdr->skip_mode_enabled && imin(bw4, bh4) > 1)

    {

        const int smctx = t->a->skip_mode[bx4] + t->l.skip_mode[by4];

        b->skip_mode = dav1d_msac_decode_bool_adapt(&ts->msac,

                           ts->cdf.m.skip_mode[smctx]);

        if (DEBUG_BLOCK_INFO)

            printf("Post-skipmode[%d]: r=%d\n", b->skip_mode, ts->msac.rng);

    } else {

        b->skip_mode = 0;

    }

    // skip

    if (b->skip_mode || (seg && seg->skip)) {

        b->skip = 1;

    } else {

        const int sctx = t->a->skip[bx4] + t->l.skip[by4];

        b->skip = dav1d_msac_decode_bool_adapt(&ts->msac, ts->cdf.m.skip[sctx]);

        if (DEBUG_BLOCK_INFO)

            printf("Post-skip[%d]: r=%d\n", b->skip, ts->msac.rng);

    }

    // segment_id

    if (f->frame_hdr->segmentation.enabled &&

        f->frame_hdr->segmentation.update_map &&

        !f->frame_hdr->segmentation.seg_data.preskip)

    {

        if (!b->skip && f->frame_hdr->segmentation.temporal &&

            (seg_pred = dav1d_msac_decode_bool_adapt(&ts->msac,

                            ts->cdf.m.seg_pred[t->a->seg_pred[bx4] +

                            t->l.seg_pred[by4]])))

        {

            // temporal predicted seg_id

            if (f->prev_segmap) {

                unsigned seg_id = get_prev_frame_segid(f, t->by, t->bx, w4, h4,

                                                       f->prev_segmap,

                                                       f->b4_stride);

                if (seg_id >= 8) return -1;

                b->seg_id = seg_id;

            } else {

                b->seg_id = 0;

            }

        } else {

            int seg_ctx;

            const unsigned pred_seg_id =

                get_cur_frame_segid(t->by, t->bx, have_top, have_left,

                                    &seg_ctx, f->cur_segmap, f->b4_stride);

            if (b->skip) {

                b->seg_id = pred_seg_id;

            } else {

                const unsigned diff = dav1d_msac_decode_symbol_adapt8(&ts->msac,

                                          ts->cdf.m.seg_id[seg_ctx],

                                          DAV1D_MAX_SEGMENTS - 1);

                const unsigned last_active_seg_id =

                    f->frame_hdr->segmentation.seg_data.last_active_segid;

                b->seg_id = neg_deinterleave(diff, pred_seg_id,

                                             last_active_seg_id + 1);

                if (b->seg_id > last_active_seg_id) b->seg_id = 0; // error?

            }

            if (b->seg_id >= DAV1D_MAX_SEGMENTS) b->seg_id = 0; // error?

        }

        seg = &f->frame_hdr->segmentation.seg_data.d[b->seg_id];

        if (DEBUG_BLOCK_INFO)

            printf("Post-segid[postskip;%d]: r=%d\n",

                   b->seg_id, ts->msac.rng);

    }

    // cdef index

    if (!b->skip) {

        const int idx = f->seq_hdr->sb128 ? ((t->bx & 16) >> 4) +

                                           ((t->by & 16) >> 3) : 0;

        if (t->cur_sb_cdef_idx_ptr[idx] == -1) {

            const int v = dav1d_msac_decode_bools(&ts->msac,

                              f->frame_hdr->cdef.n_bits);

            t->cur_sb_cdef_idx_ptr[idx] = v;

            if (bw4 > 16) t->cur_sb_cdef_idx_ptr[idx + 1] = v;

            if (bh4 > 16) t->cur_sb_cdef_idx_ptr[idx + 2] = v;

            if (bw4 == 32 && bh4 == 32) t->cur_sb_cdef_idx_ptr[idx + 3] = v;

            if (DEBUG_BLOCK_INFO)

                printf("Post-cdef_idx[%d]: r=%d\n",

                        *t->cur_sb_cdef_idx_ptr, ts->msac.rng);

        }

    }

    // delta-q/lf

    if (!((t->bx | t->by) & (31 >> !f->seq_hdr->sb128))) {

        const int prev_qidx = ts->last_qidx;

        const int have_delta_q = f->frame_hdr->delta.q.present &&

            (bs != (f->seq_hdr->sb128 ? BS_128x128 : BS_64x64) || !b->skip);

        uint32_t prev_delta_lf = ts->last_delta_lf.u32;

        if (have_delta_q) {

            int delta_q = dav1d_msac_decode_symbol_adapt4(&ts->msac,

                                                          ts->cdf.m.delta_q, 3);

            if (delta_q == 3) {

                const int n_bits = 1 + dav1d_msac_decode_bools(&ts->msac, 3);

                delta_q = dav1d_msac_decode_bools(&ts->msac, n_bits) +

                          1 + (1 << n_bits);

            }

            if (delta_q) {

                if (dav1d_msac_decode_bool_equi(&ts->msac)) delta_q = -delta_q;

                delta_q *= 1 << f->frame_hdr->delta.q.res_log2;

            }

            ts->last_qidx = iclip(ts->last_qidx + delta_q, 1, 255);

            if (have_delta_q && DEBUG_BLOCK_INFO)

                printf("Post-delta_q[%d->%d]: r=%d\n",

                       delta_q, ts->last_qidx, ts->msac.rng);

            if (f->frame_hdr->delta.lf.present) {

                const int n_lfs = f->frame_hdr->delta.lf.multi ?

                    f->cur.p.layout != DAV1D_PIXEL_LAYOUT_I400 ? 4 : 2 : 1;

                for (int i = 0; i < n_lfs; i++) {

                    int delta_lf = dav1d_msac_decode_symbol_adapt4(&ts->msac,

                        ts->cdf.m.delta_lf[i + f->frame_hdr->delta.lf.multi], 3);

                    if (delta_lf == 3) {

                        const int n_bits = 1 + dav1d_msac_decode_bools(&ts->msac, 3);

                        delta_lf = dav1d_msac_decode_bools(&ts->msac, n_bits) +

                                   1 + (1 << n_bits);

                    }

                    if (delta_lf) {

                        if (dav1d_msac_decode_bool_equi(&ts->msac))

                            delta_lf = -delta_lf;

                        delta_lf *= 1 << f->frame_hdr->delta.lf.res_log2;

                    }

                    ts->last_delta_lf.i8[i] =

                        iclip(ts->last_delta_lf.i8[i] + delta_lf, -63, 63);

                    if (have_delta_q && DEBUG_BLOCK_INFO)

                        printf("Post-delta_lf[%d:%d]: r=%d\n", i, delta_lf,

                               ts->msac.rng);

                }

            }

        }

        if (ts->last_qidx == f->frame_hdr->quant.yac) {

            // assign frame-wide q values to this sb

            ts->dq = f->dq;

        } else if (ts->last_qidx != prev_qidx) {