/*

 * Copyright (c) 2016, Alliance for Open Media. All rights reserved

 *

 * This source code is subject to the terms of the BSD 3-Clause Clear License and

 * the Alliance for Open Media Patent License 1.0. If the BSD 3-Clause Clear License

 * was not distributed with this source code in the LICENSE file, you can

 * obtain it at https://www.aomedia.org/license. If the Alliance for Open

 * Media Patent License 1.0 was not distributed with this source code in the

 * PATENTS file, you can obtain it at https://www.aomedia.org/license/patent-license.

 */

#include <stdio.h>

#include <stdlib.h>

#include <math.h>

#include <string.h>

#include "enc_cdef.h"

#include <stdint.h>

#include "aom_dsp_rtcd.h"

#include "cdef_copy.h"

#include "svt_log.h"

#include "rd_cost.h"

#include "rc_process.h"

static INLINE uint64_t mse_8xn_16bit_c(const uint16_t* src, const uint16_t* dst, const int32_t dstride,

                                       const int32_t height, uint8_t subsampling_factor) {

    uint64_t sum = 0;

    int32_t  i, j;

    for (i = 0; i < height; i += subsampling_factor) {

        for (j = 0; j < 8; j++) {

            int32_t e = dst[i * dstride + j] - src[8 * i + j];

            sum += e * e;

        }

    }

    return sum;

}

static INLINE uint64_t mse_4xn_16bit_c(const uint16_t* src, const uint16_t* dst, const int32_t dstride,

                                       const int32_t height, uint8_t subsampling_factor) {

    uint64_t sum = 0;

    int32_t  i, j;

    for (i = 0; i < height; i += subsampling_factor) {

        for (j = 0; j < 4; j++) {

            int32_t e = dst[i * dstride + j] - src[4 * i + j];

            sum += e * e;

        }

    }

    return sum;

}

static INLINE uint64_t mse_8xn_8bit_c(const uint8_t* src, const uint8_t* dst, const int32_t dstride,

                                      const int32_t height, uint8_t subsampling_factor) {

    uint64_t sum = 0;

    int32_t  i, j;

    for (i = 0; i < height; i += subsampling_factor) {

        for (j = 0; j < 8; j++) {

            int32_t e = dst[i * dstride + j] - src[8 * i + j];

            sum += e * e;

        }

    }

    return sum;

}

static INLINE uint64_t mse_4xn_8bit_c(const uint8_t* src, const uint8_t* dst, const int32_t dstride,

                                      const int32_t height, uint8_t subsampling_factor) {

    uint64_t sum = 0;

    int32_t  i, j;

    for (i = 0; i < height; i += subsampling_factor) {

        for (j = 0; j < 4; j++) {

            int32_t e = dst[i * dstride + j] - src[4 * i + j];

            sum += e * e;

        }

    }

    return sum;

}

/* Compute MSE only on the blocks we filtered. */

uint64_t svt_aom_compute_cdef_dist_16bit_c(const uint16_t* dst, int32_t dstride, const uint16_t* src,

                                           const CdefList* dlist, int32_t cdef_count, BlockSize bsize,

                                           int32_t coeff_shift, uint8_t subsampling_factor) {

    uint64_t sum = 0;

    int32_t  bi, bx, by;

    if (bsize == BLOCK_8X8) {

        for (bi = 0; bi < cdef_count; bi++) {

            by = dlist[bi].by;

            bx = dlist[bi].bx;

            sum += mse_8xn_16bit_c(

                &src[bi << (3 + 3)], &dst[(by << 3) * dstride + (bx << 3)], dstride, 8, subsampling_factor);

        }

    } else if (bsize == BLOCK_4X8) {

        for (bi = 0; bi < cdef_count; bi++) {

            by = dlist[bi].by;

            bx = dlist[bi].bx;

            sum += mse_4xn_16bit_c(

                &src[bi << (3 + 2)], &dst[(by << 3) * dstride + (bx << 2)], dstride, 8, subsampling_factor);

        }

    } else if (bsize == BLOCK_8X4) {

        for (bi = 0; bi < cdef_count; bi++) {

            by = dlist[bi].by;

            bx = dlist[bi].bx;

            sum += mse_8xn_16bit_c(

                &src[bi << (2 + 3)], &dst[(by << 2) * dstride + (bx << 3)], dstride, 4, subsampling_factor);

        }

    } else {

        assert(bsize == BLOCK_4X4);

        for (bi = 0; bi < cdef_count; bi++) {

            by = dlist[bi].by;

            bx = dlist[bi].bx;

            sum += mse_4xn_16bit_c(

                &src[bi << (2 + 2)], &dst[(by << 2) * dstride + (bx << 2)], dstride, 4, subsampling_factor);

        }

    }

    return sum >> 2 * coeff_shift;

}

uint64_t svt_aom_compute_cdef_dist_8bit_c(const uint8_t* dst8, int32_t dstride, const uint8_t* src8,

                                          const CdefList* dlist, int32_t cdef_count, BlockSize bsize,

                                          int32_t coeff_shift, uint8_t subsampling_factor) {

    uint64_t sum = 0;

    int32_t  bi, bx, by;

    if (bsize == BLOCK_8X8) {

        for (bi = 0; bi < cdef_count; bi++) {

            by = dlist[bi].by;

            bx = dlist[bi].bx;

            sum += mse_8xn_8bit_c(

                &src8[bi << (3 + 3)], &dst8[(by << 3) * dstride + (bx << 3)], dstride, 8, subsampling_factor);

        }

    } else if (bsize == BLOCK_4X8) {

        for (bi = 0; bi < cdef_count; bi++) {

            by = dlist[bi].by;

            bx = dlist[bi].bx;

            sum += mse_4xn_8bit_c(

                &src8[bi << (3 + 2)], &dst8[(by << 3) * dstride + (bx << 2)], dstride, 8, subsampling_factor);

        }

    } else if (bsize == BLOCK_8X4) {

        for (bi = 0; bi < cdef_count; bi++) {

            by = dlist[bi].by;

            bx = dlist[bi].bx;

            sum += mse_8xn_8bit_c(

                &src8[bi << (2 + 3)], &dst8[(by << 2) * dstride + (bx << 3)], dstride, 4, subsampling_factor);

        }

    } else {

        assert(bsize == BLOCK_4X4);

        for (bi = 0; bi < cdef_count; bi++) {

            by = dlist[bi].by;

            bx = dlist[bi].bx;

            sum += mse_4xn_8bit_c(

                &src8[bi << (2 + 2)], &dst8[(by << 2) * dstride + (bx << 2)], dstride, 4, subsampling_factor);

        }

    }

    return sum >> 2 * coeff_shift;

}

int32_t svt_sb_compute_cdef_list(PictureControlSet* pcs, const Av1Common* const cm, int32_t mi_row, int32_t mi_col,

                                 CdefList* dlist, BlockSize bs) {

    int32_t maxc = cm->mi_cols - mi_col;

    int32_t maxr = cm->mi_rows - mi_row;

    if (bs == BLOCK_128X128 || bs == BLOCK_128X64) {

        maxc = AOMMIN(maxc, MI_SIZE_128X128);

    } else {

        maxc = AOMMIN(maxc, MI_SIZE_64X64);

    }

    if (bs == BLOCK_128X128 || bs == BLOCK_64X128) {

        maxr = AOMMIN(maxr, MI_SIZE_128X128);

    } else {

        maxr = AOMMIN(maxr, MI_SIZE_64X64);

    }

    // CDEF lists 8x8 luma units; one unit = 2x2 mi blocks (MI_SIZE = 4px), so the mi step is 2 and

    // the unit index is the mi index >> 1. (mi_size_wide/high[BLOCK_8X8] == 2, fixed.)

    int32_t count = 0;

    for (int32_t r = 0; r < maxr; r += 2) {

        MbModeInfo** const row0 = pcs->mi_grid_base + (mi_row + r) * pcs->mi_stride + mi_col;

        MbModeInfo** const row1 = row0 + pcs->mi_stride;

        for (int32_t c = 0; c < maxc; c += 2) {

            uint8_t all_skip = row0[c]->block_mi.skip & row0[c + 1]->block_mi.skip & row1[c]->block_mi.skip &

                row1[c + 1]->block_mi.skip;

            dlist[count].by = (uint8_t)(r >> 1);

            dlist[count].bx = (uint8_t)(c >> 1);

            count += 1 - all_skip;

        }

    }

    return count;

}

/*

Loop over all 64x64 filter blocks and perform the CDEF filtering for each block, using

the filter strength pairs chosen in finish_cdef_search().

*/

#if CDEF_8BITS_PATH

// Native 8-bit src8 builder for both interior and frame-edge fbs: builds the 8-bit src8 buffer

// DIRECTLY from recon + the 8-bit line/col halo buffers (plain 8->8 copies, no widen/narrow), then

// filters all blocks via svt_cdef_filter_fb_lbd. For frame-edge fbs the off-frame halo is left as

// garbage and masked geometrically by the boundary-aware kernel (no 16-bit sentinel buffer).

// linebuf/colbuf are viewed as uint8 here (8-bit pixel halos); the persistent buffers are over-

// allocated as uint16 for HBD, but the 8-bit path only touches the uint8 half (bit depth is fixed

// per pipeline, so no path mixes element sizes within a run).

// Off-frame recon reads are avoided: top/tl/tr route to linebuf when prev_row_cdef==1 (true for the

// first fb row), left routes to colbuf for fbc==0, and the body copy clamps its bottom/right extent.

static inline void cdef_apply_fb(uint8_t* rec_buff, uint32_t rec_stride, uint8_t* src8, uint8_t* linebuf_pli,

                                 uint8_t* colbuf_pli, const uint8_t* prev_row_cdef, int fbr, int fbc, int nvfb,

                                 int nhfb, int coffset, int mhl2, int mwl2, int nhb, int nvb, int stride, int rend,

                                 int cend, int cstart, int cdef_left, uint8_t dir[CDEF_NBLOCKS][CDEF_NBLOCKS],

                                 int* dirinit, int32_t var[CDEF_NBLOCKS][CDEF_NBLOCKS], int pli, CdefList* dlist,

                                 int cdef_count, int cdef_strength, int damping, int coeff_shift, int xdec, int ydec,

                                 int frame_top, int frame_left, int frame_bottom, int frame_right) {

    const int vsz = nvb << mhl2;

    const int hsz = nhb << mwl2;

    // CDEF taps reach only +-CDEF_HALO px, so the recon copies need only a CDEF_HALO-wide halo even

    // though the buffer is padded to HBORDER/VBORDER for aligned SIMD loads (the extra border is

    // never read into any output). Name the halo geometry inside the padded buffer.

    const int halo_col = CDEF_HBORDER - CDEF_HALO; // left-halo column

    const int halo_row = CDEF_VBORDER - CDEF_HALO; // top-halo row

    const int top_off  = halo_row * CDEF_BSTRIDE + CDEF_HBORDER; // src8 (top halo, body col 0)

    const int tl_off   = halo_row * CDEF_BSTRIDE + halo_col; // src8 (top halo, left halo col)

    const int tr_off   = halo_row * CDEF_BSTRIDE + CDEF_HBORDER + hsz; // src8 (top halo, right body edge)

    const int lrow     = halo_row * stride; // linebuf row offset of the halo

    const int cs       = cstart < -CDEF_HALO ? -CDEF_HALO : cstart;

    // Off-frame bottom/right have no in-frame halo: clamp the recon body copy so it never reads past

    // the frame edge (those taps are masked geometrically by the bounded kernel).

    const int body_rows = frame_bottom ? vsz : vsz + CDEF_HALO;

    const int cright    = frame_right ? hsz : cend - (CDEF_HBORDER - CDEF_HALO);

    // Body + top halo in ONE copy from recon (also brings the left/right and bottom halo). The top

    // halo is overwritten from the pre-filter linebuf below wherever the top neighbour fb was already

    // CDEF-filtered (its current recon is post-filter and unusable as halo).

    svt_cdef_copy_rect8(&src8[(CDEF_VBORDER - CDEF_HALO) * CDEF_BSTRIDE + CDEF_HBORDER + cs],

                        CDEF_BSTRIDE,

                        rec_buff,

                        (MI_SIZE_64X64 << mhl2) * fbr - CDEF_HALO,

                        coffset + cs,

                        rec_stride,

                        CDEF_HALO + body_rows,

                        cright - cs);

    // Overwrite top center / top-left / top-right from the pre-filter linebuf where that neighbour

    // was filtered (else the recon copied above is correct).

    if (prev_row_cdef[fbc]) {

        svt_cdef_copy_rect8(&src8[top_off], CDEF_BSTRIDE, &linebuf_pli[lrow + coffset], 0, 0, stride, CDEF_HALO, hsz);

    }

    if (prev_row_cdef[fbc - 1]) {

        svt_cdef_copy_rect8(

            &src8[tl_off], CDEF_BSTRIDE, &linebuf_pli[lrow + coffset - CDEF_HALO], 0, 0, stride, CDEF_HALO, CDEF_HALO);

    }

    if (prev_row_cdef[fbc + 1]) {

        svt_cdef_copy_rect8(

            &src8[tr_off], CDEF_BSTRIDE, &linebuf_pli[lrow + coffset + hsz], 0, 0, stride, CDEF_HALO, CDEF_HALO);

    }

    // Left halo (from colbuf, if the left fb was filtered).

    if (cdef_left) {

        svt_cdef_copy_rect8(

            &src8[halo_col], CDEF_BSTRIDE, &colbuf_pli[halo_col], 0, 0, CDEF_HBORDER, rend + CDEF_VBORDER, CDEF_HALO);

    }

    // Save right halo edge to colbuf for the fb to the right.

    if (fbc < nhfb - 1) {

        svt_cdef_copy_rect8(&colbuf_pli[halo_col],

                            CDEF_HBORDER,

                            &src8[hsz + halo_col],

                            0,

                            0,

                            CDEF_BSTRIDE,

                            rend + CDEF_VBORDER,

                            CDEF_HALO);

    }

    // Save bottom edge to linebuf for the fb below.

    if (fbr < nvfb - 1) {

        svt_cdef_copy_rect8(&linebuf_pli[coffset],

                            stride,

                            &rec_buff[((MI_SIZE_64X64 << mhl2) * (fbr + 1) - CDEF_VBORDER) * rec_stride + coffset],

                            0,

                            0,

                            rec_stride,

                            CDEF_VBORDER,

                            hsz);

    }

    if (cdef_strength || !*dirinit) {

        svt_cdef_filter_fb_lbd(&rec_buff[rec_stride * (MI_SIZE_64X64 * fbr << mhl2) + (fbc * MI_SIZE_64X64 << mwl2)],

                               rec_stride,

                               &src8[CDEF_VBORDER * CDEF_BSTRIDE + CDEF_HBORDER],

                               frame_top,

                               frame_left,

                               frame_bottom,

                               frame_right,

                               vsz,

                               hsz,

                               xdec,

                               ydec,

                               dir,

                               dirinit,

                               var,

                               pli,

                               dlist,

                               cdef_count,

                               cdef_strength,

                               damping,

                               coeff_shift,

                               1);

    }

}

#endif

void svt_av1_cdef_frame(SequenceControlSet* scs, PictureControlSet* pcs) {

    PictureParentControlSet* ppcs    = pcs->ppcs;

    Av1Common*               cm      = ppcs->av1_cm;

    FrameHeader*             frm_hdr = &ppcs->frm_hdr;

#if CONFIG_ENABLE_HIGH_BIT_DEPTH

    const bool is_16bit = scs->is_16bit_pipeline;

#else

    const bool is_16bit = false;

#endif

    EbPictureBufferDesc* recon_pic;

    svt_aom_get_recon_pic(pcs, &recon_pic, is_16bit);

    const int32_t num_planes = av1_num_planes(&scs->seq_header.color_config);

#if !CDEF_8BITS_PATH || CONFIG_ENABLE_HIGH_BIT_DEPTH

    DECLARE_ALIGNED(16, uint16_t, src[CDEF_INBUF_SIZE]); // 16-bit sentinel buffer (HBD / non-8bit path only)

#endif

#if CDEF_8BITS_PATH

    // 8-bit mirror of `src` for the native interior filter (no 16-bit-lane work).

    DECLARE_ALIGNED(16, uint8_t, src8[CDEF_INBUF_SIZE]);

#endif

    uint16_t*      linebuf[3];

    uint16_t*      colbuf[3];

    CdefList       dlist_local[MI_SIZE_64X64 * MI_SIZE_64X64];

    CdefList*      dlist;

    uint8_t *      row_cdef, *prev_row_cdef, *curr_row_cdef;

    int32_t        cdef_count;

    const uint32_t sb_size = scs->super_block_size;

    // Reuse the dlist/count computed by the search (SB=64 and the frame was actually searched).

    const bool use_dlist_cache = sb_size == 64 && !ppcs->cdef_search_ctrls.use_reference_cdef_fs &&

        !ppcs->cdef_search_ctrls.use_qp_strength;

    int32_t        mi_wide_l2[3];

    int32_t        mi_high_l2[3];

    int32_t        xdec[3];

    int32_t        ydec[3];

    int32_t        coeff_shift = AOMMAX(scs->static_config.encoder_bit_depth - 8, 0);

    const int32_t  nvfb        = (cm->mi_rows + MI_SIZE_64X64 - 1) / MI_SIZE_64X64;

    const int32_t  nhfb        = (cm->mi_cols + MI_SIZE_64X64 - 1) / MI_SIZE_64X64;

    const uint32_t cdef_size   = sizeof(*row_cdef) * (nhfb + 2) * 2;

    // Persistent scratch: (re)allocate only when it must grow, reuse across frames.

    if (pcs->cdef_row_cdef_sz < cdef_size) {

        svt_aom_free(pcs->cdef_row_cdef);

        pcs->cdef_row_cdef    = (uint8_t*)svt_aom_malloc(cdef_size);

        pcs->cdef_row_cdef_sz = cdef_size;

    }

    row_cdef = pcs->cdef_row_cdef;

    assert(row_cdef != NULL);

    memset(row_cdef, 1, cdef_size);

    prev_row_cdef = row_cdef + 1;

    curr_row_cdef = prev_row_cdef + nhfb + 2;

    for (int32_t pli = 0; pli < num_planes; pli++) {

        int32_t subsampling_x = (pli == 0) ? 0 : 1;

        int32_t subsampling_y = (pli == 0) ? 0 : 1;

        xdec[pli]             = subsampling_x; //CHKN xd->plane[pli].subsampling_x;

        ydec[pli]             = subsampling_y; //CHKN  xd->plane[pli].subsampling_y;

        mi_wide_l2[pli]       = MI_SIZE_LOG2 - subsampling_x; //CHKN xd->plane[pli].subsampling_x;

        mi_high_l2[pli]       = MI_SIZE_LOG2 - subsampling_y; //CHKN xd->plane[pli].subsampling_y;

    }

    const int32_t stride = (cm->mi_cols << MI_SIZE_LOG2) + 2 * CDEF_HBORDER;

    for (int32_t pli = 0; pli < num_planes; pli++) {

        const uint32_t lb_sz = sizeof(*linebuf) * CDEF_VBORDER * stride;

        const uint32_t cb_sz = sizeof(*colbuf) * ((CDEF_BLOCKSIZE << mi_high_l2[pli]) + 2 * CDEF_VBORDER) *

            CDEF_HBORDER;

        if (pcs->cdef_linebuf_sz[pli] < lb_sz) {

            svt_aom_free(pcs->cdef_linebuf[pli]);

            pcs->cdef_linebuf[pli]    = (uint16_t*)svt_aom_malloc(lb_sz);

            pcs->cdef_linebuf_sz[pli] = lb_sz;

        }

        if (pcs->cdef_colbuf_sz[pli] < cb_sz) {

            svt_aom_free(pcs->cdef_colbuf[pli]);

            pcs->cdef_colbuf[pli]    = (uint16_t*)svt_aom_malloc(cb_sz);

            pcs->cdef_colbuf_sz[pli] = cb_sz;

        }

        linebuf[pli] = pcs->cdef_linebuf[pli];

        colbuf[pli]  = pcs->cdef_colbuf[pli];

    }

    // Frame-level check: if every UV strength entry is 0, no chroma block

    // will ever be filtered.  In that case skip all chroma border copies

    // (including linebuf/colbuf saves) for the entire frame

    bool chroma_filter_off = (num_planes > 1);

    if (chroma_filter_off) {

        for (int32_t i = 0; i < ppcs->nb_cdef_strengths; i++) {

            if (frm_hdr->cdef_params.cdef_uv_strength[i] != 0) {

                chroma_filter_off = false;

                break;

            }

        }

    }

    const int32_t active_planes      = chroma_filter_off ? 1 : num_planes;

    int32_t       mbmi_cdef_strength = 0;

    for (int32_t fbr = 0; fbr < nvfb; fbr++) {

        int32_t cdef_left = 1;

        for (int32_t fbc = 0; fbc < nhfb; fbc++) {

            // per-fb state read by the row below; must be written even for skipped fbs

            curr_row_cdef[fbc] = 0;

            const MbModeInfo* mbmi = pcs->mi_grid_base[MI_SIZE_64X64 * fbr * cm->mi_stride + MI_SIZE_64X64 * fbc];

            // Find the index of the CDEF strength for the filter block. With a single frame strength

            // (nb_cdef_strengths==1, e.g. non-searched frames) every SB uses index 0; only multi-strength

            // (searched) frames carry a per-SB index, so the per-SB read is only needed then.

            if (ppcs->nb_cdef_strengths > 1) {

                mbmi_cdef_strength = mbmi->cdef_strength;

            }

            uint8_t cdef_y_strength  = frm_hdr->cdef_params.cdef_y_strength[mbmi_cdef_strength];

            uint8_t cdef_uv_strength = frm_hdr->cdef_params.cdef_uv_strength[mbmi_cdef_strength];

            if (cdef_y_strength == 0 && cdef_uv_strength == 0) {

                cdef_left = 0;

                continue;

            }

            if (use_dlist_cache) {

                cdef_count = pcs->cdef_fb_list[fbr * nhfb + fbc].cdef_count;

                dlist      = pcs->cdef_fb_list[fbr * nhfb + fbc].dlist;

            } else {

                cdef_count = svt_sb_compute_cdef_list(

                    pcs, cm, fbr * MI_SIZE_64X64, fbc * MI_SIZE_64X64, dlist_local, BLOCK_64X64);

                dlist = dlist_local;

            }

            if (cdef_count == 0) {

                cdef_left = 0;

                continue;

            }

            curr_row_cdef[fbc] = 1;

            int32_t nhb, nvb;

            int32_t cstart = 0;

            assert(mbmi != NULL && "CDEF ERROR: Skipping Current FB");

            assert((ppcs->nb_cdef_strengths == 1 || mbmi->cdef_strength != -1) && "CDEF ERROR: Skipping Current FB");

            if (!cdef_left) {

                //CHKN if the left block has not been filtered, then we can use samples on the left as input.

                cstart = -CDEF_HBORDER;

            }

            nhb = AOMMIN(MI_SIZE_64X64, cm->mi_cols - MI_SIZE_64X64 * fbc);

            nvb = AOMMIN(MI_SIZE_64X64, cm->mi_rows - MI_SIZE_64X64 * fbr);

            // A fb edge is a frame boundary iff the fb sits in the first/last fb row or column.

            // nvfb/nhfb = ceil(mi_rows|mi_cols / 64), so only the last fb row/col can reach the frame

            // edge -> the per-mi cm->mi_* compare the original used is always equivalent to this.

            const int32_t frame_top    = (fbr == 0);

            const int32_t frame_left   = (fbc == 0);

            const int32_t frame_bottom = (fbr == nvfb - 1);

            const int32_t frame_right  = (fbc == nhfb - 1);

            int dirinit = !(ppcs->cdef_search_ctrls.use_reference_cdef_fs || ppcs->cdef_search_ctrls.use_qp_strength);

            // When SB 128 is used, the search for certain blocks is skipped, so dir/var info is not generated

            // In those cases, must generate info here

            if (sb_size == 128) {

                const BlockSize bsize = mbmi->bsize;

                if (((fbc & 1) && (bsize == BLOCK_128X128 || bsize == BLOCK_128X64)) ||

                    ((fbr & 1) && (bsize == BLOCK_128X128 || bsize == BLOCK_64X128))) {

                    dirinit = 0;

                }

            }

            uint8_t (*dir)[CDEF_NBLOCKS][CDEF_NBLOCKS] = &pcs->cdef_dir_data[fbr * nhfb + fbc].dir;

            int32_t (*var)[CDEF_NBLOCKS][CDEF_NBLOCKS] = &pcs->cdef_dir_data[fbr * nhfb + fbc].var;

            for (int32_t pli = 0; pli < active_planes; pli++) {

                // Strength index for this plane: luma uses Y, chroma planes both use UV. The packed

                // value is passed to the leaf filters, which decode primary/secondary internally.

                const int32_t cdef_strength = pli ? cdef_uv_strength : cdef_y_strength;

                const int32_t damping       = frm_hdr->cdef_params.cdef_damping;

                int32_t hsize = nhb << mi_wide_l2[pli];

                int32_t vsize = nvb << mi_high_l2[pli];

                int32_t rend  = (fbr == nvfb - 1) ? vsize : vsize + CDEF_VBORDER;

                int32_t cend  = (fbc == nhfb - 1) ? hsize : hsize + CDEF_HBORDER;

                int32_t  coffset    = fbc * MI_SIZE_64X64 << mi_wide_l2[pli];

                EbByte   rec_buff   = recon_pic->buffer[pli];

                uint32_t rec_stride = recon_pic->stride[pli];

#if CDEF_8BITS_PATH

                // 8-bit pipeline: build src8 directly from recon (interior AND frame-edge fbs) and

                // filter via the boundary-aware hybrid -- no 16-bit sentinel buffer, no widen/narrow.

                if (!is_16bit) {

                    cdef_apply_fb(rec_buff,

                                  rec_stride,

                                  src8,

                                  (uint8_t*)linebuf[pli],

                                  (uint8_t*)colbuf[pli],

                                  prev_row_cdef,

                                  fbr,

                                  fbc,

                                  nvfb,

                                  nhfb,

                                  coffset,

                                  mi_high_l2[pli],

                                  mi_wide_l2[pli],

                                  nhb,

                                  nvb,

                                  stride,

                                  rend,

                                  cend,

                                  cstart,

                                  cdef_left,

                                  *dir,

                                  &dirinit,

                                  *var,

                                  pli,

                                  dlist,

                                  cdef_count,

                                  cdef_strength,

                                  damping,

                                  coeff_shift,

                                  xdec[pli],

                                  ydec[pli],

                                  frame_top,

                                  frame_left,

                                  frame_bottom,

                                  frame_right);

                    continue;

                }

#endif

#if !CDEF_8BITS_PATH || CONFIG_ENABLE_HIGH_BIT_DEPTH

                /* Copy in the pixels we need from the current superblock for

                   deringing. CDEF taps reach only +-CDEF_HALO px, so a 2-px halo into the

                   padded src buffer suffices (ring + interior blocks both read only +-2). */

                const int halo_col = CDEF_HBORDER - CDEF_HALO; // left/right halo column

                const int halo_row = CDEF_VBORDER - CDEF_HALO; // top halo row

                const int top_off  = halo_row * CDEF_BSTRIDE + CDEF_HBORDER; // src (top halo, body col 0)

                const int tl_off   = halo_row * CDEF_BSTRIDE + halo_col; // src (top halo, left halo col)

                const int tr_off   = halo_row * CDEF_BSTRIDE + CDEF_HBORDER + hsize; // src (top halo, right body edge)

                const int lrow     = halo_row * stride; // linebuf row offset of the halo

                const int e_cs     = cstart < -CDEF_HALO ? -CDEF_HALO : cstart;

                const int e_cright = cend > hsize + CDEF_HALO ? hsize + CDEF_HALO : cend;

                const int e_rbot   = rend > vsize + CDEF_HALO ? vsize + CDEF_HALO : rend;

                svt_aom_copy_sb8_16(&src[CDEF_VBORDER * CDEF_BSTRIDE + CDEF_HBORDER + e_cs],

                                    CDEF_BSTRIDE,

                                    rec_buff,

                                    (MI_SIZE_64X64 << mi_high_l2[pli]) * fbr,

                                    coffset + e_cs,

                                    rec_stride,

                                    e_rbot,

                                    e_cright - e_cs,

                                    is_16bit);

                if (!prev_row_cdef[fbc]) {

                    svt_aom_copy_sb8_16(&src[top_off],

                                        CDEF_BSTRIDE,

                                        rec_buff,

                                        (MI_SIZE_64X64 << mi_high_l2[pli]) * fbr - CDEF_HALO,

                                        coffset,

                                        rec_stride,

                                        CDEF_HALO,

                                        hsize,

                                        is_16bit);

                } else if (fbr > 0) {

                    svt_aom_copy_rect(

                        &src[top_off], CDEF_BSTRIDE, &linebuf[pli][lrow + coffset], stride, CDEF_HALO, hsize);

                } else {

                    svt_aom_fill_rect(&src[top_off], CDEF_BSTRIDE, CDEF_HALO, hsize, CDEF_VERY_LARGE);

                }

                if (!prev_row_cdef[fbc - 1]) {

                    svt_aom_copy_sb8_16(&src[tl_off],

                                        CDEF_BSTRIDE,

                                        rec_buff,

                                        (MI_SIZE_64X64 << mi_high_l2[pli]) * fbr - CDEF_HALO,

                                        coffset - CDEF_HALO,

                                        rec_stride,

                                        CDEF_HALO,

                                        CDEF_HALO,

                                        is_16bit);

                } else if (fbr > 0 && fbc > 0) {

                    svt_aom_copy_rect(&src[tl_off],

                                      CDEF_BSTRIDE,

                                      &linebuf[pli][lrow + coffset - CDEF_HALO],

                                      stride,

                                      CDEF_HALO,

                                      CDEF_HALO);

                } else {

                    svt_aom_fill_rect(&src[tl_off], CDEF_BSTRIDE, CDEF_HALO, CDEF_HALO, CDEF_VERY_LARGE);

                }

                if (!prev_row_cdef[fbc + 1]) {

                    svt_aom_copy_sb8_16(&src[tr_off],

                                        CDEF_BSTRIDE,

                                        rec_buff,

                                        (MI_SIZE_64X64 << mi_high_l2[pli]) * fbr - CDEF_HALO,

                                        coffset + hsize,

                                        rec_stride,

                                        CDEF_HALO,

                                        CDEF_HALO,

                                        is_16bit);

                } else if (fbr > 0 && fbc < nhfb - 1) {

                    svt_aom_copy_rect(&src[tr_off],

                                      CDEF_BSTRIDE,

                                      &linebuf[pli][lrow + coffset + hsize],

                                      stride,

                                      CDEF_HALO,

                                      CDEF_HALO);

                } else {

                    svt_aom_fill_rect(&src[tr_off], CDEF_BSTRIDE, CDEF_HALO, CDEF_HALO, CDEF_VERY_LARGE);

                }

                if (cdef_left) {

                    /* If we deringed the superblock on the left then we need to copy in

                       saved pixels. */

                    svt_aom_copy_rect(&src[halo_col],

                                      CDEF_BSTRIDE,

                                      &colbuf[pli][halo_col],

                                      CDEF_HBORDER,

                                      rend + CDEF_VBORDER,

                                      CDEF_HALO);

                }

                /* Saving pixels in case we need to dering the superblock on the

                    right. */

                if (fbc < nhfb - 1) {

                    svt_aom_copy_rect(&colbuf[pli][halo_col],

                                      CDEF_HBORDER,

                                      &src[hsize + halo_col],

                                      CDEF_BSTRIDE,

                                      rend + CDEF_VBORDER,

                                      CDEF_HALO);

                }

                if (fbr < nvfb - 1) {

                    svt_aom_copy_sb8_16(&linebuf[pli][coffset],

                                        stride,

                                        rec_buff,

                                        (MI_SIZE_64X64 << mi_high_l2[pli]) * (fbr + 1) - CDEF_VBORDER,

                                        coffset,

                                        rec_stride,

                                        CDEF_VBORDER,

                                        hsize,

                                        is_16bit);

                }

                if (frame_top) {

                    svt_aom_fill_rect(&src[halo_row * CDEF_BSTRIDE + halo_col],

                                      CDEF_BSTRIDE,

                                      CDEF_HALO,

                                      hsize + 2 * CDEF_HALO,

                                      CDEF_VERY_LARGE);

                }

                if (frame_left) {

                    svt_aom_fill_rect(&src[halo_row * CDEF_BSTRIDE + halo_col],

                                      CDEF_BSTRIDE,

                                      vsize + 2 * CDEF_HALO,

                                      CDEF_HALO,

                                      CDEF_VERY_LARGE);

                }

                if (frame_bottom) {

                    svt_aom_fill_rect(&src[(vsize + CDEF_VBORDER) * CDEF_BSTRIDE + halo_col],

                                      CDEF_BSTRIDE,

                                      CDEF_HALO,

                                      hsize + 2 * CDEF_HALO,

                                      CDEF_VERY_LARGE);

                }

                if (frame_right) {

                    svt_aom_fill_rect(&src[halo_row * CDEF_BSTRIDE + CDEF_HBORDER + hsize],

                                      CDEF_BSTRIDE,

                                      vsize + 2 * CDEF_HALO,

                                      CDEF_HALO,

                                      CDEF_VERY_LARGE);

                }

                // if ppcs->cdef_ctrls.use_reference_cdef_fs is true, then search was not performed

                // Therefore, need to make sure dir and var are initialized

                if (cdef_strength || !dirinit) {

                    // 8-bit content is handled above by cdef_apply_fb (which continues); this

                    // 16-bit sentinel path now serves only the HBD (is_16bit) pipeline (and the

                    // whole-frame fallback when CDEF_8BITS_PATH is disabled).

                    svt_cdef_filter_fb(

                        is_16bit ? NULL

                                 : &rec_buff[rec_stride * (MI_SIZE_64X64 * fbr << mi_high_l2[pli]) +

                                             (fbc * MI_SIZE_64X64 << mi_wide_l2[pli])],

                        is_16bit ? &((uint16_t*)rec_buff)[rec_stride * (MI_SIZE_64X64 * fbr << mi_high_l2[pli]) +

                                                          (fbc * MI_SIZE_64X64 << mi_wide_l2[pli])]

                                 : NULL,

                        rec_stride,

                        &src[CDEF_VBORDER * CDEF_BSTRIDE + CDEF_HBORDER],

                        xdec[pli],

                        ydec[pli],

                        *dir,

                        &dirinit,

                        *var,

                        pli,

                        dlist,

                        cdef_count,

                        cdef_strength,

                        damping,

                        coeff_shift,

                        1); // no subsampling

                }

#endif

            }

            cdef_left = 1; //CHKN filtered data is written back directy to recFrame.

        }

        {

            uint8_t* tmp  = prev_row_cdef;

            prev_row_cdef = curr_row_cdef;

            curr_row_cdef = tmp;

        }

    }

}

///-------search

/*

 * Search for the best luma+chroma strength to add as an option, knowing we

 * already selected nb_strengths options

 *

 * Params:

 *

 * lev0 : Array of indices of selected luma strengths.

 * lev1 : Array of indices of selected chroma strengths.

 * nb_strengths : Number of selected (Luma_strength, Chroma_strength) pairs.

 * mse : Array of luma and chroma filtering mse values.

 * sb_count : Number of filter blocks in the frame.

 * start_gi : starting strength index for the search of the additional strengths.

 * end_gi : End index for the for the search of the additional strengths.

*/

uint64_t svt_search_one_dual_c(int* lev0, int* lev1, int nb_strengths, uint64_t** mse[2], int sb_count, int start_gi,

                               int end_gi) {

    uint64_t      tot_mse[TOTAL_STRENGTHS][TOTAL_STRENGTHS];

    int32_t       i, j;

    uint64_t      best_tot_mse    = (uint64_t)1 << 63;

    int32_t       best_id0        = 0;

    int32_t       best_id1        = 0;

    const int32_t total_strengths = end_gi;

    memset(tot_mse, 0, sizeof(tot_mse));

    /* Loop over the filter blocks in the frame */

    for (i = 0; i < sb_count; i++) {

        int32_t  gi;

        uint64_t best_mse = (uint64_t)1 << 63;

        /* Loop over the already selected nb_strengths (Luma_strength,

           Chroma_strength) pairs, and find the pair that has the smallest mse

           (best_mse) for the current filter block.*/

        /* Find best mse among already selected options. */

        for (gi = 0; gi < nb_strengths; gi++) {

            uint64_t curr = mse[0][i][lev0[gi]];

            curr += mse[1][i][lev1[gi]];

            if (curr < best_mse) {

                best_mse = curr;

            }

        }

        /* Loop over the set of available (Luma_strength, Chroma_strength)

           pairs, identify any that provide an mse better than best_mse from the

           step above for the current filter block, and update any corresponding

           total mse (tot_mse[j][k]). */

        /* Find best mse when adding each possible new option. */

        for (j = start_gi; j < total_strengths; j++) {

            int32_t k;

            for (k = start_gi; k < total_strengths; k++) {

                uint64_t best = best_mse;

                uint64_t curr = mse[0][i][j];

                curr += mse[1][i][k];

                if (curr < best) {

                    best = curr;

                }

                tot_mse[j][k] += best;

            }

        }

    }

    /* Loop over the additionally searched (Luma_strength, Chroma_strength) pairs

       from the step above, and identify any such pair that provided the best mse for

       the whole frame. The identified pair would be added to the set of already selected pairs. */

    for (j = start_gi; j < total_strengths; j++) { // Loop over the additionally searched luma strengths

        int32_t k;

        for (k = start_gi; k < total_strengths; k++) { // Loop over the additionally searched chroma strengths

            if (tot_mse[j][k] < best_tot_mse) {

                best_tot_mse = tot_mse[j][k];

                best_id0     = j; // index for the best luma strength

                best_id1     = k; // index for the best chroma strength

            }

        }

    }

    lev0[nb_strengths] = best_id0; // Add the identified luma strength to the list of selected luma strengths

    lev1[nb_strengths] = best_id1; // Add the identified chroma strength to the list of selected chroma strengths

    return best_tot_mse;

}

/*

 * Search for the set of luma+chroma strengths that minimizes mse.

 *

 * Params:

 *

 * best_lev0 : Array of indices of selected luma strengths.

 * best_lev1 : Array of indices of selected chroma strengths.

 * nb_strengths : Number of selected (Luma_strength, Chroma_strength) pairs.

 * mse : Array of luma and chroma filtering mse values.

 * sb_count : Number of filter blocks in the frame.

 * start_gi : starting strength index for the search of the additional strengths.

 * end_gi : End index for the for the search of the additional strengths.

*/

static uint64_t joint_strength_search_dual(int32_t* best_lev0, int32_t* best_lev1, int32_t nb_strengths,

                                           uint64_t** mse[2], int32_t sb_count, int32_t start_gi, int32_t end_gi) {

    uint64_t best_tot_mse;

    int32_t  i;

    best_tot_mse = (uint64_t)1 << 63;

    /* Greedy search: add one strength options at a time.

    Determine nb_strengths (Luma_strength, Chroma_strength) pairs.

    The list of nb_strengths pairs is determined by adding one such pair at

    a time through the call to the function search_one_dual. When the

    function search_one_dual is called, the search accounts for the

    strength pairs that have already been added in the previous iteration of

    the loop below. The loop below returns in the end best_tot_mse

    representing the best filtering mse for the whole frame based on the

    selected list of best (Luma_strength, Chroma_strength) pairs.

    */

    for (i = 0; i < nb_strengths; i++) {

        best_tot_mse = svt_search_one_dual(best_lev0, best_lev1, i, mse, sb_count, start_gi, end_gi);

    }

    /* Performing further refinements on the search based on the results

    from the step above. Trying to refine the greedy search by reconsidering each

    already-selected option. */

    for (i = 0; i < 4 * nb_strengths; i++) {

        int32_t j;

        for (j = 0; j < nb_strengths - 1; j++) {

            best_lev0[j] = best_lev0[j + 1];

            best_lev1[j] = best_lev1[j + 1];

        }

        best_tot_mse = svt_search_one_dual(best_lev0, best_lev1, nb_strengths - 1, mse, sb_count, start_gi, end_gi);

    }

    return best_tot_mse;

}

// This kernel is ported/adapted from libaom (AV1 reference implementation).

// Original logic inspired by aom_pick_cdef_from_qp().

// Adjusted to match SVT-AV1 data structures and pipeline integration.

static void svt_pick_cdef_from_qp(PictureParentControlSet* ppcs, int32_t is_screen_content, int32_t* pred_y_strength,

                                  int32_t* pred_uv_strength) {

    FrameHeader*  frm_hdr    = &ppcs->frm_hdr;

    const uint8_t bit_depth  = ppcs->enhanced_pic->bit_depth;

    const int32_t base_q_idx = frm_hdr->quantization_params.base_q_idx;

    int32_t q = svt_aom_ac_quant_qtx(base_q_idx, 0, bit_depth);

    q >>= (bit_depth - 8);

    int32_t y_f1 = 0, y_f2 = 0;

    int32_t uv_f1 = 0, uv_f2 = 0;

    const int32_t is_intra = (frm_hdr->frame_type == KEY_FRAME || frm_hdr->frame_type == INTRA_ONLY_FRAME);

    if (is_screen_content) {

        y_f1 = (int32_t)(5.88217781e-06 * q * q + 6.10391455e-03 * q + 9.95043102e-02);

        y_f2 = (int32_t)(-7.79934857e-06 * q * q + 6.58957830e-03 * q + 8.81045025e-01);

        uv_f1 = (int32_t)(-6.79500136e-06 * q * q + 1.02695586e-02 * q + 1.36126802e-01);

        uv_f2 = (int32_t)(-9.99613695e-08 * q * q - 1.79361339e-05 * q + 1.17022324e+0);

    } else if (!is_intra) {

        y_f1 = (int32_t)roundf(q * q * -0.0000023593946f + q * 0.0068615186f + 0.02709886f);

        y_f2 = (int32_t)roundf(q * q * -0.00000057629734f + q * 0.0013993345f + 0.03831067f);

        uv_f1 = (int32_t)roundf(q * q * -0.0000007095069f + q * 0.0034628846f + 0.00887099f);

        uv_f2 = (int32_t)roundf(q * q * 0.00000023874085f + q * 0.00028223585f + 0.05576307f);

    } else { // Intra

        y_f1 = (int32_t)roundf(q * q * 0.0000033731974f + q * 0.008070594f + 0.0187634f);

        y_f2 = (int32_t)roundf(q * q * 0.0000029167343f + q * 0.0027798624f + 0.0079405f);

        uv_f1 = (int32_t)roundf(q * q * -0.0000130790995f + q * 0.012892405f - 0.00748388f);

        uv_f2 = (int32_t)roundf(q * q * 0.0000032651783f + q * 0.00035520183f + 0.00228092f);

    }

    // Clamp to AV1 limits

    y_f1  = clamp(y_f1, 0, 15);

    y_f2  = clamp(y_f2, 0, 3);

    uv_f1 = clamp(uv_f1, 0, 15);

    uv_f2 = clamp(uv_f2, 0, 3);

    // Pack primary + secondary

    *pred_y_strength  = y_f1 * CDEF_SEC_STRENGTHS + y_f2;

    *pred_uv_strength = uv_f1 * CDEF_SEC_STRENGTHS + uv_f2;

}

// Propagate cdef_strength to all 64x64 mi

static INLINE void propagate_cdef_strength(PictureControlSet* pcs, int32_t sb_index, int8_t strength) {

    MbModeInfo* mbmi    = pcs->mi_grid_base[sb_index];

    mbmi->cdef_strength = strength;

    switch (mbmi->bsize) {

    case BLOCK_128X128:

        pcs->mi_grid_base[sb_index + MI_SIZE_64X64]->cdef_strength                                  = strength;

        pcs->mi_grid_base[sb_index + MI_SIZE_64X64 * pcs->mi_stride]->cdef_strength                 = strength;

        pcs->mi_grid_base[sb_index + MI_SIZE_64X64 * pcs->mi_stride + MI_SIZE_64X64]->cdef_strength = strength;

        break;

    case BLOCK_128X64: