#include "alf.h" #include #include #include #include #include #include "cabac.h" #include "rdo.h" #include "strategies/strategies-alf.h" #include "kvz_math.h" #include "reshape.h" extern kvz_pixel kvz_fast_clip_32bit_to_pixel(int32_t value); #if MAX_NUM_CC_ALF_FILTERS>1 typedef struct filter_idx_count { uint64_t count; uint8_t filter_idx; } filter_idx_count; static int comparator(const void *v1, const void *v2) { const filter_idx_count *p1 = (filter_idx_count *)v1; const filter_idx_count *p2 = (filter_idx_count *)v2; return (p1->count < p2->count); } #endif static void reset_alf_covariance(alf_covariance *alf, int num_bins) { if (num_bins > 0) { alf->num_bins = num_bins; } alf->pix_acc = 0; memset(alf->y, 0, sizeof(alf->y)); memset(alf->ee, 0, sizeof(alf->ee)); } void kvz_reset_cc_alf_aps_param(cc_alf_filter_param *cc_alf) { memset(cc_alf->cc_alf_filter_enabled, false, sizeof(cc_alf->cc_alf_filter_enabled)); memset(cc_alf->cc_alf_filter_idx_enabled, false, sizeof(cc_alf->cc_alf_filter_idx_enabled)); memset(cc_alf->cc_alf_coeff, 0, sizeof(cc_alf->cc_alf_coeff)); cc_alf->cc_alf_filter_count[0] = cc_alf->cc_alf_filter_count[1] = MAX_NUM_CC_ALF_FILTERS; cc_alf->number_valid_components = 3; cc_alf->new_cc_alf_filter[0] = cc_alf->new_cc_alf_filter[1] = 0; } static void reset_alf_param(alf_aps *src) { memset(src->enabled_flag, false, sizeof(src->enabled_flag)); memset(src->non_linear_flag, false, sizeof(src->non_linear_flag)); memset(src->luma_coeff, 0, sizeof(src->luma_coeff)); memset(src->luma_clipp, 0, sizeof(src->luma_clipp)); src->num_alternatives_chroma = 1; memset(src->chroma_coeff, 0, sizeof(src->chroma_coeff)); memset(src->chroma_clipp, 0, sizeof(src->chroma_clipp)); memset(src->filter_coeff_delta_idx, 0, sizeof(src->filter_coeff_delta_idx)); memset(src->alf_luma_coeff_flag, true, sizeof(src->alf_luma_coeff_flag)); src->num_luma_filters = 1; src->alf_luma_coeff_delta_flag = false; memset(src->new_filter_flag, 0, sizeof(src->new_filter_flag)); } static void reset_aps(alf_aps *src, bool cc_alf_enabled) { src->aps_type = 0; src->aps_id = -1; src->temporal_id = 0; src->layer_id = 0; reset_alf_param(src); if (cc_alf_enabled) { kvz_reset_cc_alf_aps_param(&src->cc_alf_aps_param); } } void kvz_set_aps_map(videoframe_t* frame, enum kvz_alf alf_type) { frame->alf_param_set_map = malloc(ALF_CTB_MAX_NUM_APS * sizeof(param_set_map)); for (int aps_idx = 0; aps_idx < ALF_CTB_MAX_NUM_APS; aps_idx++) { frame->alf_param_set_map[aps_idx + T_ALF_APS].b_changed = false; reset_aps(&frame->alf_param_set_map[aps_idx + T_ALF_APS].parameter_set, alf_type == KVZ_ALF_FULL); } } static void init_ctu_alternative_chroma(const alf_aps *alf_param, uint8_t* ctu_alts[MAX_NUM_COMPONENT], const int32_t num_ctus) { uint8_t alt_idx = 0; for (int ctu_idx = 0; ctu_idx < num_ctus; ++ctu_idx) { ctu_alts[COMPONENT_Cb][ctu_idx] = alt_idx; if ((ctu_idx + 1) * alf_param->num_alternatives_chroma >= (alt_idx + 1) * num_ctus) ++alt_idx; } alt_idx = 0; for (int ctu_idx = 0; ctu_idx < num_ctus; ++ctu_idx) { ctu_alts[COMPONENT_Cr][ctu_idx] = alt_idx; if ((ctu_idx + 1) * alf_param->num_alternatives_chroma >= (alt_idx + 1) * num_ctus) ++alt_idx; } } static int16_t clip_alf(const int16_t clip, const int16_t ref, const int16_t val0, const int16_t val1) { return CLIP(-clip, +clip, val0 - ref) + CLIP(-clip, +clip, val1 - ref); } static void get_clip_max(const alf_covariance *cov, int *clip_max) { const int num_coeff = cov->num_coeff; for (int k = 0; k < num_coeff - 1; ++k) { clip_max[k] = 0; bool inc = true; while (inc && clip_max[k] + 1 < cov->num_bins && cov->y[clip_max[k] + 1][k] == cov->y[clip_max[k]][k]) { for (int l = 0; inc && l < num_coeff; ++l) { if (cov->ee[clip_max[k]][0][k][l] != cov->ee[clip_max[k] + 1][0][k][l]) { inc = false; } } if (inc) { ++clip_max[k]; } } } clip_max[num_coeff - 1] = 0; } static void reduce_clip_cost(const alf_covariance *cov, int *clip) { for (int k = 0; k < cov->num_coeff - 1; ++k) { bool dec = true; while (dec && clip[k] > 0 && cov->y[clip[k] - 1][k] == cov->y[clip[k]][k]) { for (int l = 0; dec && l < cov->num_coeff; ++l) { if (cov->ee[clip[k]][clip[l]][k][l] != cov->ee[clip[k] - 1][clip[l]][k][l]) { dec = false; } } if (dec) { --clip[k]; } } } } static void set_ey_from_clip(const alf_covariance *cov, const int* clip, double ee[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF], double y[MAX_NUM_ALF_LUMA_COEFF], int size) { for (int k = 0; k < size; k++) { y[k] = cov->y[clip[k]][k]; for (int l = 0; l < size; l++) { ee[k][l] = cov->ee[clip[k]][clip[l]][k][l]; } } } static int gns_cholesky_dec(double inp_matr[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF], double out_matr[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF], int num_eq) { double inv_diag[MAX_NUM_ALF_LUMA_COEFF]; /* Vector of the inverse of diagonal entries of outMatr */ for (int i = 0; i < num_eq; i++) { for (int j = i; j < num_eq; j++) { /* Compute the scaling factor */ double scale = inp_matr[i][j]; if (i > 0) { for (int k = i - 1; k >= 0; k--) { scale -= out_matr[k][j] * out_matr[k][i]; } } /* Compute i'th row of outMatr */ if (i == j) { if (scale <= 0.0000001) // if(scale <= 0 ) /* If inpMatr is singular */ { return 0; } else /* Normal operation */ { inv_diag[i] = 1.0 / (out_matr[i][i] = sqrt(scale)); } } else { out_matr[i][j] = scale * inv_diag[i]; /* Upper triangular part */ out_matr[j][i] = 0.0; /* Lower triangular part set to 0 */ } } } return 1; /* Signal that Cholesky factorization is successfully performed */ } static void gns_transpose_backsubstitution(double u[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF], double* rhs, double* x, int order) { /* Backsubstitution starts */ x[0] = rhs[0] / u[0][0]; /* First row of U' */ for (int i = 1; i < order; i++) { /* For the rows 1..order-1 */ double sum = 0; //Holds backsubstitution from already handled rows for (int j = 0; j < i; j++) /* Backsubst already solved unknowns */ { sum += x[j] * u[j][i]; } x[i] = (rhs[i] - sum) / u[i][i]; /* i'th component of solution vect. */ } } static void gns_backsubstitution(double r[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF], double* z, int size, double* a) { size--; a[size] = z[size] / r[size][size]; for (int i = size - 1; i >= 0; i--) { double sum = 0; for (int j = i + 1; j <= size; j++) { sum += r[i][j] * a[j]; } a[i] = (z[i] - sum) / r[i][i]; } } static int gns_solve_by_chol(double lhs[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF], double rhs[MAX_NUM_ALF_LUMA_COEFF], double *x, int num_eq) { double aux[MAX_NUM_ALF_LUMA_COEFF]; /* Auxiliary vector */ double u[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF]; /* Upper triangular Cholesky factor of lhs */ int res = 1; // Signal that Cholesky factorization is successfully performed /* The equation to be solved is LHSx = rhs */ /* Compute upper triangular U such that U'*U = lhs */ if (gns_cholesky_dec(lhs, u, num_eq)) /* If Cholesky decomposition has been successful */ { /* Now, the equation is U'*U*x = rhs, where U is upper triangular * Solve U'*aux = rhs for aux */ gns_transpose_backsubstitution(u, rhs, aux, num_eq); /* The equation is now U*x = aux, solve it for x (new motion coefficients) */ gns_backsubstitution(u, aux, num_eq, x); } else /* lhs was singular */ { res = 0; /* Regularize lhs */ for (int i = 0; i < num_eq; i++) { lhs[i][i] += 0.0001; } /* Compute upper triangular U such that U'*U = regularized lhs */ res = gns_cholesky_dec(lhs, u, num_eq); if (!res) { memset(x, 0, sizeof(double)*num_eq); return 0; } /* Solve U'*aux = rhs for aux */ gns_transpose_backsubstitution(u, rhs, aux, num_eq); /* Solve U*x = aux for x */ gns_backsubstitution(u, aux, num_eq, x); } return res; } static int gns_solve_by_chol_clip_gns(alf_covariance *cov, const int *clip, double *x, int num_eq) { double lhs[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF]; double rhs[MAX_NUM_ALF_LUMA_COEFF]; set_ey_from_clip(cov, clip, lhs, rhs, num_eq); return gns_solve_by_chol(lhs, rhs, x, num_eq); } static double calculate_error(const alf_covariance *cov, const int *clip, const double *coeff) { double sum = 0; for (int i = 0; i < cov->num_coeff; i++) { sum += coeff[i] * cov->y[clip[i]][i]; } return cov->pix_acc - sum; } static double optimize_filter(const alf_covariance *cov, int* clip, double *f, bool optimize_clip) { const int size = cov->num_coeff; int clip_max[MAX_NUM_ALF_LUMA_COEFF]; double err_best, err_last; double ke[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF]; double ky[MAX_NUM_ALF_LUMA_COEFF]; if (optimize_clip) { // Start by looking for min clipping that has no impact => max_clipping get_clip_max(cov, clip_max); for (int k = 0; k < size; ++k) { clip[k] = MAX(clip_max[k], clip[k]); clip[k] = MIN(clip[k], cov->num_bins - 1); } } set_ey_from_clip(cov, clip, ke, ky, size); gns_solve_by_chol(ke, ky, f, size); err_best = calculate_error(cov, clip, f); int step = optimize_clip ? (cov->num_bins + 1) / 2 : 0; while (step > 0) { double err_min = err_best; int idx_min = -1; int inc_min = 0; for (int k = 0; k < size - 1; ++k) { if (clip[k] - step >= clip_max[k]) { clip[k] -= step; ky[k] = cov->y[clip[k]][k]; for (int l = 0; l < size; l++) { ke[k][l] = cov->ee[clip[k]][clip[l]][k][l]; ke[l][k] = cov->ee[clip[l]][clip[k]][l][k]; } gns_solve_by_chol(ke, ky, f, size); err_last = calculate_error(cov, clip, f); if (err_last < err_min) { err_min = err_last; idx_min = k; inc_min = -step; } clip[k] += step; } if (clip[k] + step < cov->num_bins) { clip[k] += step; ky[k] = cov->y[clip[k]][k]; for (int l = 0; l < size; l++) { ke[k][l] = cov->ee[clip[k]][clip[l]][k][l]; ke[l][k] = cov->ee[clip[l]][clip[k]][l][k]; } gns_solve_by_chol(ke, ky, f, size); err_last = calculate_error(cov, clip, f); if (err_last < err_min) { err_min = err_last; idx_min = k; inc_min = step; } clip[k] -= step; } ky[k] = cov->y[clip[k]][k]; for (int l = 0; l < size; l++) { ke[k][l] = cov->ee[clip[k]][clip[l]][k][l]; ke[l][k] = cov->ee[clip[l]][clip[k]][l][k]; } } if (idx_min >= 0) { err_best = err_min; clip[idx_min] += inc_min; ky[idx_min] = cov->y[clip[idx_min]][idx_min]; for (int l = 0; l < size; l++) { ke[idx_min][l] = cov->ee[clip[idx_min]][clip[l]][idx_min][l]; ke[l][idx_min] = cov->ee[clip[l]][clip[idx_min]][l][idx_min]; } } else { --step; } } if (optimize_clip) { // test all max for (int k = 0; k < size - 1; ++k) { clip_max[k] = 0; } double ke_max[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF]; double ky_max[MAX_NUM_ALF_LUMA_COEFF]; set_ey_from_clip(cov, clip_max, ke_max, ky_max, size); gns_solve_by_chol(ke_max, ky_max, f, size); err_last = calculate_error(cov, clip_max, f); if (err_last < err_best) { err_best = err_last; for (int k = 0; k < size; ++k) { clip[k] = clip_max[k]; } } else { // update clip to reduce coding cost reduce_clip_cost(cov, clip); // update f with best solution gns_solve_by_chol(ke, ky, f, size); } } return err_best; } static double optimize_filter_clip(alf_covariance *cov, int* clip) { double f[MAX_NUM_ALF_LUMA_COEFF]; return optimize_filter(cov, clip, f, true); } static double optimize_filter_gns_calc(alf_covariance *cov, const int* clip, double *f, int size) { gns_solve_by_chol_clip_gns(cov, clip, f, size); return calculate_error(cov, clip, f); } static double calc_error_for_coeffs(const alf_covariance *cov, const int *clip, const int *coeff, const int num_coeff, const int bit_depth) { double factor = 1 << (bit_depth - 1); double error = 0; for (int i = 0; i < num_coeff; i++) //diagonal { double sum = 0; for (int j = i + 1; j < num_coeff; j++) { sum += cov->ee[clip[i]][clip[j]][i][j] * coeff[j]; } error += ((cov->ee[clip[i]][clip[i]][i][i] * coeff[i] + sum * 2) / factor - 2 * cov->y[clip[i]][i]) * coeff[i]; } return error / factor; } static double calc_error_for_cc_alf_coeffs(const alf_covariance *cov, const int16_t* coeff, const int num_coeff, const int bit_depth) { double factor = 1 << (bit_depth - 1); double error = 0; for (int i = 0; i < num_coeff; i++) // diagonal { double sum = 0; for (int j = i + 1; j < num_coeff; j++) { // E[j][i] = E[i][j], sum will be multiplied by 2 later sum += cov->ee[0][0][i][j] * coeff[j]; } error += ((cov->ee[0][0][i][i] * coeff[i] + sum * 2) / factor - 2 * cov->y[0][i]) * coeff[i]; } return error / factor; } static int length_uvlc(int ui_code) { int ui_length = 1; int ui_temp = ++ui_code; assert(ui_temp); // "Integer overflow" while (1 != ui_temp) { ui_temp >>= 1; ui_length += 2; } // Take care of cases where ui_length > 32 return (ui_length >> 1) + ((ui_length + 1) >> 1); } static double get_dist_coeff_force_0(bool* coded_var_bins, double error_force_0_coeff_tab[MAX_NUM_ALF_CLASSES][2], int* bits_var_bin, int zero_bits_var_bin, const int num_filters, double lambda) { double dist_force_0 = 0; memset(coded_var_bins, 0, sizeof(*coded_var_bins) * MAX_NUM_ALF_CLASSES); for (int filt_idx = 0; filt_idx < num_filters; filt_idx++) { double cost_diff = (error_force_0_coeff_tab[filt_idx][0] + lambda * zero_bits_var_bin) - (error_force_0_coeff_tab[filt_idx][1] + lambda * bits_var_bin[filt_idx]); coded_var_bins[filt_idx] = cost_diff > 0 ? true : false; dist_force_0 += error_force_0_coeff_tab[filt_idx][coded_var_bins[filt_idx] ? 1 : 0]; } return dist_force_0; } static double get_dist_force_0(const alf_aps *alf_param, channel_type channel, const int num_filters, double error_tab_force_0_coeff[MAX_NUM_ALF_CLASSES][2], bool* coded_var_bins, double lambda, int filter_coeff_set[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF], int filter_clipp_set[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF] ) { int num_coeff = channel == CHANNEL_TYPE_LUMA ? 13 : 7; int bits_var_bin[MAX_NUM_ALF_CLASSES]; for (int ind = 0; ind < num_filters; ++ind) { bits_var_bin[ind] = 0; for (int i = 0; i < num_coeff - 1; i++) { bits_var_bin[ind] += length_uvlc(abs(filter_coeff_set[ind][i])); if (abs(filter_coeff_set[ind][i]) != 0) bits_var_bin[ind] += 1; } } int zero_bits_var_bin = 0; for (int i = 0; i < num_coeff - 1; i++) { zero_bits_var_bin += length_uvlc(0); } if (alf_param->non_linear_flag[CHANNEL_TYPE_LUMA]) { for (int ind = 0; ind < num_filters; ++ind) { for (int i = 0; i < num_coeff - 1; i++) { if (!abs(filter_coeff_set[ind][i])) { filter_clipp_set[ind][i] = 0; } } } } double dist_force_0 = get_dist_coeff_force_0(coded_var_bins, error_tab_force_0_coeff, bits_var_bin, zero_bits_var_bin, num_filters, lambda); return dist_force_0; } static int get_cost_filter_coeff_force_0(const alf_aps *alf_param, channel_type channel, const int num_filters, bool* coded_var_bins, int p_diff_q_filter_coeff_int_pp[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF], int filter_clipp_set[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF] ) { const int num_coeff = channel == CHANNEL_TYPE_LUMA ? 13 : 7; int len = 0; // Filter coefficients for (int ind = 0; ind < num_filters; ++ind) { if (coded_var_bins[ind]) { for (int i = 0; i < num_coeff - 1; i++) { len += length_uvlc(abs(p_diff_q_filter_coeff_int_pp[ind][i])); // alf_coeff_luma_delta[i][j] if ((abs(p_diff_q_filter_coeff_int_pp[ind][i]) != 0)) len += 1; } } else { for (int i = 0; i < num_coeff - 1; i++) { len += length_uvlc(0); // alf_coeff_luma_delta[i][j] } } } if (alf_param->non_linear_flag[CHANNEL_TYPE_LUMA]) { for (int ind = 0; ind < num_filters; ++ind) { for (int i = 0; i < num_coeff - 1; i++) { if (!abs(p_diff_q_filter_coeff_int_pp[ind][i])) { filter_clipp_set[ind][i] = 0; } len += 2; } } } return len; } static int length_filter_coeffs(channel_type channel, const int num_filters, int filter_coeff[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF]) { int num_coeff = channel == CHANNEL_TYPE_LUMA ? 13 : 7; int bit_cnt = 0; for (int ind = 0; ind < num_filters; ++ind) { for (int i = 0; i < num_coeff - 1; i++) { bit_cnt += length_uvlc(abs(filter_coeff[ind][i])); if (abs(filter_coeff[ind][i]) != 0) bit_cnt += 1; } } return bit_cnt; } static int get_cost_filter_coeff(channel_type channel, const int num_filters, int p_diff_q_filter_coeff_int_pp[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF]) { return length_filter_coeffs(channel, num_filters, p_diff_q_filter_coeff_int_pp); // alf_coeff_luma_delta[i][j]; } static int get_cost_filter_clipp(channel_type channel, const int num_filters, int p_diff_q_filter_coeff_int_pp[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF], int filter_clipp_set[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF]) { int num_coeff = channel == CHANNEL_TYPE_LUMA ? 13 : 7; for (int filter_idx = 0; filter_idx < num_filters; ++filter_idx) { for (int i = 0; i < num_coeff - 1; i++) { if (!abs(p_diff_q_filter_coeff_int_pp[filter_idx][i])) { filter_clipp_set[filter_idx][i] = 0; } } } return (num_filters * (num_coeff - 1)) << 1; } static int get_non_filter_coeff_rate(alf_aps *aps) { int len = 0 // alf_coefficients_delta_flag + 2 // slice_alf_chroma_idc u(2) + length_uvlc(aps->num_luma_filters - 1); // alf_luma_num_filters_signalled_minus1 ue(v) if (aps->num_luma_filters > 1) { const int coeff_length = kvz_math_ceil_log2(aps->num_luma_filters); //#if JVET_O0491_HLS_CLEANUP for (int i = 0; i < MAX_NUM_ALF_CLASSES; i++) { len += coeff_length; } } return len; } static double calculate_error_opt_filt(alf_covariance *cov, const int *clip) { double c[MAX_NUM_ALF_LUMA_COEFF]; return optimize_filter_gns_calc(cov, clip, c, cov->num_coeff); } static int get_chroma_coeff_rate(alf_aps* aps, int alt_idx) { int i_bits = 0; const int num_coeff = 7; // Filter coefficients for (int i = 0; i < num_coeff - 1; i++) { i_bits += length_uvlc(abs(aps->chroma_coeff[alt_idx][i])); // alf_coeff_chroma[alt_idx][i] if ((aps->chroma_coeff[alt_idx][i]) != 0) i_bits += 1; } if (aps->non_linear_flag[CHANNEL_TYPE_CHROMA]) { for (int i = 0; i < num_coeff - 1; i++) { if (!abs(aps->chroma_coeff[alt_idx][i])) { aps->chroma_clipp[alt_idx][i] = 0; } } i_bits += ((num_coeff - 1) << 1); } return i_bits; } static double get_filtered_distortion(alf_covariance* cov, array_variables *arr_vars, const int num_classes, const int num_filters_minus1, const int num_coeff, const int bit_depth) { double dist = 0; for (int class_idx = 0; class_idx < num_classes; class_idx++) { dist += calc_error_for_coeffs(&cov[class_idx], arr_vars->filter_clipp_set[class_idx], arr_vars->filter_coeff_set[class_idx], num_coeff, bit_depth); } return dist; } static double get_unfiltered_distortion_cov_classes(alf_covariance* cov, const int num_classes) { double dist = 0; for (int class_idx = 0; class_idx < num_classes; class_idx++) { dist += cov[class_idx].pix_acc; } return dist; } static double get_unfiltered_distortion_cov_channel(alf_covariance* cov, channel_type channel) { double dist = 0; if (channel == CHANNEL_TYPE_LUMA) { dist = get_unfiltered_distortion_cov_classes(cov, MAX_NUM_ALF_CLASSES); } else { dist = get_unfiltered_distortion_cov_classes(cov, 1); } return dist; } static void add_alf_cov(alf_covariance *dst, alf_covariance *src) { int num_bins = dst->num_bins; int num_coeff = dst->num_coeff; for (int b0 = 0; b0 < num_bins; b0++) { for (int b1 = 0; b1 < num_bins; b1++) { for (int j = 0; j < num_coeff; j++) { for (int i = 0; i < num_coeff; i++) { dst->ee[b0][b1][j][i] += src->ee[b0][b1][j][i]; } } } } for (int b = 0; b < num_bins; b++) { for (int j = 0; j < num_coeff; j++) { dst->y[b][j] += src->y[b][j]; } } dst->pix_acc += src->pix_acc; } static void add_alf_cov_lhs_rhs(alf_covariance *dst, alf_covariance *lhs, alf_covariance *rhs) { int num_coeff = lhs->num_coeff; int num_bins = lhs->num_bins; for (int b0 = 0; b0 < num_bins; b0++) { for (int b1 = 0; b1 < num_bins; b1++) { for (int j = 0; j < num_coeff; j++) { for (int i = 0; i < num_coeff; i++) { dst->ee[b0][b1][j][i] = lhs->ee[b0][b1][j][i] + rhs->ee[b0][b1][j][i]; } } } } for (int b = 0; b < num_bins; b++) { for (int j = 0; j < num_coeff; j++) { dst->y[b][j] = lhs->y[b][j] + rhs->y[b][j]; } } dst->pix_acc = lhs->pix_acc + rhs->pix_acc; } static void get_frame_stat(alf_covariance* frame_cov, alf_covariance* ctb_cov, bool* ctb_enable_flags, uint8_t* ctb_alt_idx, const int num_classes, int alt_idx, const int32_t num_ctus) { const channel_type channel = (!ctb_alt_idx ? CHANNEL_TYPE_LUMA : CHANNEL_TYPE_CHROMA); bool is_luma = channel == CHANNEL_TYPE_LUMA ? true : false; for (int ctu_idx = 0; ctu_idx < num_ctus; ctu_idx++) { if (ctb_enable_flags[ctu_idx]) { for (int class_idx = 0; class_idx < num_classes; class_idx++) { if (is_luma || alt_idx == ctb_alt_idx[ctu_idx]) { add_alf_cov(&frame_cov[is_luma ? class_idx : alt_idx], &ctb_cov[(ctu_idx * num_classes) + class_idx]); } } } } } static void get_frame_stats(alf_info_t *alf_info, channel_type channel, const int32_t num_ctus) { bool is_luma = channel == CHANNEL_TYPE_LUMA ? true : false; int num_classes = is_luma ? MAX_NUM_ALF_CLASSES : 1; int num_alternatives = is_luma ? 1 : alf_info->alf_param_temp.num_alternatives_chroma; // When calling this function m_ctuEnableFlag shall be set to 0 for CTUs using alternative APS // Here we compute frame stats for building new alternative filters for (int alt_idx = 0; alt_idx < num_alternatives; ++alt_idx) { for (int i = 0; i < num_classes; i++) { is_luma ? reset_alf_covariance(&alf_info->alf_covariance_frame_luma[i], MAX_ALF_NUM_CLIPPING_VALUES) : reset_alf_covariance(&alf_info->alf_covariance_frame_chroma[alt_idx], MAX_ALF_NUM_CLIPPING_VALUES); } if (is_luma) { get_frame_stat(alf_info->alf_covariance_frame_luma, alf_info->alf_covariance_y, alf_info->ctu_enable_flag[COMPONENT_Y], NULL, num_classes, alt_idx, num_ctus); } else { get_frame_stat(alf_info->alf_covariance_frame_chroma, alf_info->alf_covariance_u, alf_info->ctu_enable_flag[COMPONENT_Cb], alf_info->ctu_alternative[COMPONENT_Cb], num_classes, alt_idx, num_ctus); get_frame_stat(alf_info->alf_covariance_frame_chroma, alf_info->alf_covariance_v, alf_info->ctu_enable_flag[COMPONENT_Cr], alf_info->ctu_alternative[COMPONENT_Cr], num_classes, alt_idx, num_ctus); } } } static void copy_cov(alf_covariance *dst, alf_covariance *src) { dst->num_coeff = src->num_coeff; dst->num_bins = src->num_bins; memcpy(&dst->ee, &src->ee, sizeof(dst->ee)); memcpy(&dst->y, &src->y, sizeof(dst->y)); dst->pix_acc = src->pix_acc; } static void copy_alf_param(alf_aps *dst, alf_aps *src) { memcpy(dst->enabled_flag, src->enabled_flag, sizeof(dst->enabled_flag)); memcpy(dst->non_linear_flag, src->non_linear_flag, sizeof(dst->non_linear_flag)); memcpy(dst->luma_coeff, src->luma_coeff, sizeof(dst->luma_coeff)); memcpy(dst->luma_clipp, src->luma_clipp, sizeof(dst->luma_clipp)); dst->num_alternatives_chroma = src->num_alternatives_chroma; memcpy(dst->chroma_coeff, src->chroma_coeff, sizeof(dst->chroma_coeff)); memcpy(dst->chroma_clipp, src->chroma_clipp, sizeof(dst->chroma_clipp)); memcpy(dst->filter_coeff_delta_idx, src->filter_coeff_delta_idx, sizeof(dst->filter_coeff_delta_idx)); memcpy(dst->alf_luma_coeff_flag, src->alf_luma_coeff_flag, sizeof(dst->alf_luma_coeff_flag)); dst->num_luma_filters = src->num_luma_filters; dst->alf_luma_coeff_delta_flag = src->alf_luma_coeff_delta_flag; memcpy(dst->new_filter_flag, src->new_filter_flag, sizeof(dst->new_filter_flag)); } static void copy_cc_alf_param(cc_alf_filter_param *dst, cc_alf_filter_param *src) { memcpy(dst->cc_alf_filter_enabled, src->cc_alf_filter_enabled, sizeof(dst->cc_alf_filter_enabled)); memcpy(dst->cc_alf_filter_idx_enabled, src->cc_alf_filter_idx_enabled, sizeof(dst->cc_alf_filter_idx_enabled)); memcpy(dst->cc_alf_filter_count, src->cc_alf_filter_count, sizeof(dst->cc_alf_filter_count)); memcpy(dst->cc_alf_coeff, src->cc_alf_coeff, sizeof(dst->cc_alf_coeff)); memcpy(dst->new_cc_alf_filter, src->new_cc_alf_filter, sizeof(dst->new_cc_alf_filter)); dst->number_valid_components = src->number_valid_components; } static void copy_alf_param_w_channel(alf_aps* dst, alf_aps* src, channel_type channel) { if (channel == CHANNEL_TYPE_LUMA) { copy_alf_param(dst, src); } else { dst->enabled_flag[COMPONENT_Cb] = src->enabled_flag[COMPONENT_Cb]; dst->enabled_flag[COMPONENT_Cr] = src->enabled_flag[COMPONENT_Cr]; dst->num_alternatives_chroma = src->num_alternatives_chroma; dst->non_linear_flag[CHANNEL_TYPE_CHROMA] = src->non_linear_flag[CHANNEL_TYPE_CHROMA]; memcpy(dst->chroma_coeff, src->chroma_coeff, sizeof(dst->chroma_coeff)); memcpy(dst->chroma_clipp, src->chroma_clipp, sizeof(dst->chroma_clipp)); } } static void copy_aps(alf_aps *dst, alf_aps *src, bool cc_alf_enabled) { dst->aps_id = src->aps_id; dst->temporal_id = src->temporal_id; dst->layer_id = src->layer_id; dst->aps_type = src->aps_type; copy_alf_param(dst, src); if (cc_alf_enabled) { copy_cc_alf_param(&dst->cc_alf_aps_param, &src->cc_alf_aps_param); } } static void copy_aps_to_map(param_set_map *dst, alf_aps *src, int8_t aps_id, bool cc_alf_enabled) { assert(0 <= aps_id && aps_id < ALF_CTB_MAX_NUM_APS); bool found = false; for (int id = 0; id < ALF_CTB_MAX_NUM_APS; id++) { if (dst[aps_id + T_ALF_APS].parameter_set.aps_id == id) { found = true; } } if (!found) { dst[aps_id + T_ALF_APS].b_changed = true; //apsMap[apsId].p_nalu_data = 0; dst[aps_id + T_ALF_APS].parameter_set.aps_id = aps_id; dst[aps_id + T_ALF_APS].parameter_set.temporal_id = src->temporal_id; dst[aps_id + T_ALF_APS].parameter_set.layer_id = src->layer_id; dst[aps_id + T_ALF_APS].parameter_set.aps_type = src->aps_type; copy_alf_param(&dst[aps_id + T_ALF_APS].parameter_set, src); if (cc_alf_enabled) { copy_cc_alf_param(&dst[aps_id + T_ALF_APS].parameter_set.cc_alf_aps_param, &src->cc_alf_aps_param); } } } static void init_alf_covariance(alf_covariance *alf, int num_coeffs) { alf->num_coeff = num_coeffs; alf->num_bins = MAX_ALF_NUM_CLIPPING_VALUES; alf->pix_acc = 0; memset(alf->y, 0, sizeof(alf->y)); memset(alf->ee, 0, sizeof(alf->ee)); } static void copy_pixels(kvz_pixel *src, int x_src_start, int y_src_start, int src_stride, kvz_pixel *dst, int x_dst_start, int y_dst_start, int dst_stride, int width, int height) { for (int y = 0; y < height; y++) { int src_y = y_src_start + y; int dst_y = y_dst_start + y; for (int x = 0; x < width; x++) { int src_x = x_src_start + x; int dst_x = x_dst_start + x; dst[dst_y*dst_stride + dst_x] = src[src_y*src_stride + src_x]; } } } static void adjust_pixels(kvz_pixel *src, int x_start, int x_end, int y_start, int y_end, int stride, int pic_width, int pic_height) { assert(x_start <= x_end); assert(y_start <= y_end); assert(x_end <= pic_width); assert(y_end <= pic_height); //not on any edge if (x_start != 0 && y_start != 0 && x_end != pic_width && y_end != pic_height) { return; } bool top_left = (x_start == 0 && y_start == 0); bool top_right = (x_end == pic_width && y_start == 0); bool bottom_left = (x_start == 0 && y_end == pic_height); bool bottom_right = (x_end == pic_width && y_end == pic_height); //left side if (x_start == 0) { for (int y = y_start; y < y_end; y++) { src[y * stride - 4] = src[y * stride - 3] = src[y * stride - 2] = src[y * stride - 1] = src[y * stride]; } } //right side if (x_end == pic_width) { const int x_px = x_end - 1; for (int y = y_start; y < y_end; y++) { src[y * stride + x_px + 4] = src[y * stride + x_px + 3] = src[y * stride + x_px + 2] = src[y * stride + x_px + 1] = src[y * stride + x_px]; } } //top if (y_start == 0) { for (int x = x_start; x < x_end; x++) { src[-4 * stride + x] = src[-3 * stride + x] = src[-2 * stride + x] = src[-1 * stride + x] = src[x]; } } //bottom if (y_end == pic_height) { const int y_px = y_end - 1; for (int x = x_start; x < x_end; x++) { src[x + stride * (4 + y_px)] = src[x + stride * (3 + y_px)] = src[x + stride * (2 + y_px)] = src[x + stride * (1 + y_px)] = src[x + stride * y_px]; } } //left top corner if (top_left) { for (int x = -4; x < 0; x++) { src[-4 * stride + x] = src[-3 * stride + x] = src[-2 * stride + x] = src[-1 * stride + x] = src[0]; } } //right top corner if (top_right) { const int x_px = x_end - 1; for (int x = pic_width; x < pic_width + 4; x++) { src[-4 * stride + x] = src[-3 * stride + x] = src[-2 * stride + x] = src[-1 * stride + x] = src[x_px]; } } //left or right bottom corner if (bottom_left) { const int y_px = y_end - 1; for (int x = -4; x < 0; x++) { src[(4 + y_px) * stride + x] = src[(3 + y_px) * stride + x] = src[(2 + y_px) * stride + x] = src[(1 + y_px) * stride + x] = src[stride * y_px]; } } if (bottom_right) { const int x_px = x_end - 1; const int y_px = y_end - 1; for (int x = x_end; x < x_end + 4; x++) { src[(4 + y_px) * stride + x] = src[(3 + y_px) * stride + x] = src[(2 + y_px) * stride + x] = src[(1 + y_px) * stride + x] = src[stride * y_px + x_px]; } } } static void adjust_pixels_chroma(kvz_pixel *src, int x_start, int x_end, int y_start, int y_end, int stride, int pic_width, int pic_height) { assert(x_start <= x_end); assert(y_start <= y_end); assert(x_end <= pic_width); assert(y_end <= pic_height); //not on any edge if (x_start != 0 && y_start != 0 && x_end != pic_width && y_end != pic_height) { return; } bool top_left = (x_start == 0 && y_start == 0); bool top_right = (x_end == pic_width && y_start == 0); bool bottom_left = (x_start == 0 && y_end == pic_height); bool bottom_right = (x_end == pic_width && y_end == pic_height); //left side if (x_start == 0) { for (int y = y_start; y < y_end; y++) { src[y * stride - 2] = src[y * stride - 1] = src[y * stride]; } } //right side if (x_end == pic_width) { const int x_px = x_end - 1; for (int y = y_start; y < y_end; y++) { src[y * stride + x_px + 2] = src[y * stride + x_px + 1] = src[y * stride + x_px]; } } //top if (y_start == 0) { for (int x = x_start; x < x_end; x++) { src[-2 * stride + x] = src[-1 * stride + x] = src[x]; } } //bottom if (y_end == pic_height) { const int y_px = y_end - 1; for (int x = x_start; x < x_end; x++) { src[x + stride * (2 + y_px)] = src[x + stride * (1 + y_px)] = src[x + stride * y_px]; } } //left top corner if (top_left) { for (int x = -2; x < 0; x++) { src[-2 * stride + x] = src[-1 * stride + x] = src[0]; } } //right top corner if (top_right) { const int x_px = x_end - 1; for (int x = pic_width; x < pic_width + 2; x++) { src[-2 * stride + x] = src[-1 * stride + x] = src[x_px]; } } //left bottom corner if (bottom_left) { const int y_px = y_end - 1; for (int x = -2; x < 0; x++) { src[(2 + y_px) * stride + x] = src[(1 + y_px) * stride + x] = src[stride * y_px]; } } //right bottom corner if (bottom_right) { const int x_px = x_end - 1; const int y_px = y_end - 1; for (int x = x_end; x < x_end + 2; x++) { src[(2 + y_px) * stride + x] = src[(1 + y_px) * stride + x] = src[stride * y_px + x_px]; } } } static void set_ctu_enable_flag(bool **flags, channel_type channel, uint8_t value, const int32_t num_ctus) { if (channel == CHANNEL_TYPE_LUMA) { memset(flags[COMPONENT_Y], value, sizeof(bool) * num_ctus); } else { memset(flags[COMPONENT_Cr], value, sizeof(bool) * num_ctus); memset(flags[COMPONENT_Cb], value, sizeof(bool) * num_ctus); } } static void copy_ctu_enable_flag(bool **flags_dst, bool **flags_src, channel_type channel, const int32_t num_ctus) { if (channel == CHANNEL_TYPE_LUMA) { memcpy(flags_dst[COMPONENT_Y], flags_src[COMPONENT_Y], sizeof(bool) * num_ctus); } else { memcpy(flags_dst[COMPONENT_Cr], flags_src[COMPONENT_Cr], sizeof(bool) * num_ctus); memcpy(flags_dst[COMPONENT_Cb], flags_src[COMPONENT_Cb], sizeof(bool) * num_ctus); } } //-------------------------cabac writer functions------------------------ static void alf_cabac_reset_bits(cabac_data_t * const data) { data->low = 0; data->bits_left = 23; data->num_buffered_bytes = 0; data->buffered_byte = 0xff; } static void code_alf_ctu_enable_flag(encoder_state_t * const state, cabac_data_t * const cabac, uint32_t ctu_rs_addr, alf_component_id component_id, alf_aps *aps) { const encoder_control_t * const encoder = state->encoder_control; const bool alf_component_enabled = (aps != NULL) ? aps->enabled_flag[component_id] : state->slice->alf->tile_group_alf_enabled_flag[component_id]; if (encoder->cfg.alf_type && alf_component_enabled) { int frame_width_in_ctus = state->tile->frame->width_in_lcu; bool left_avail = ctu_rs_addr % frame_width_in_ctus ? 1 : 0; bool above_avail = ctu_rs_addr/ frame_width_in_ctus ? 1 : 0; int left_ctu_addr = left_avail ? ctu_rs_addr - 1 : -1; int above_ctu_addr = above_avail ? ctu_rs_addr - frame_width_in_ctus : -1; bool* ctb_alf_flag = state->tile->frame->alf_info->ctu_enable_flag[component_id]; int ctx = 0; ctx += left_ctu_addr > -1 ? (ctb_alf_flag[left_ctu_addr] ? 1 : 0) : 0; ctx += above_ctu_addr > -1 ? (ctb_alf_flag[above_ctu_addr] ? 1 : 0) : 0; cabac->cur_ctx = &(cabac->ctx.alf_ctb_flag_model[component_id * 3 + ctx]); CABAC_BIN(cabac, ctb_alf_flag[ctu_rs_addr], "alf_ctb_flag"); } } static void code_alf_ctu_enable_flags_component(encoder_state_t * const state, cabac_data_t * const cabac, alf_component_id component_id, alf_aps *aps) { const int32_t num_ctus_in_pic = state->tile->frame->width_in_lcu * state->tile->frame->height_in_lcu; for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { code_alf_ctu_enable_flag(state, cabac, ctu_idx, component_id, aps); } } static void code_alf_ctu_enable_flags_channel(encoder_state_t * const state, cabac_data_t * const cabac, channel_type channel, alf_aps *aps) { if (channel == CHANNEL_TYPE_LUMA) { if (aps->enabled_flag[COMPONENT_Y]) code_alf_ctu_enable_flags_component(state, cabac, COMPONENT_Y, aps); } else { if (aps->enabled_flag[COMPONENT_Cb]) code_alf_ctu_enable_flags_component(state, cabac, COMPONENT_Cb, aps); if (aps->enabled_flag[COMPONENT_Cr]) code_alf_ctu_enable_flags_component(state, cabac, COMPONENT_Cr, aps); } } static void code_alf_ctu_filter_index(encoder_state_t * const state, cabac_data_t * const cabac, uint32_t ctu_rs_addr, bool alf_enable_luma) { const encoder_control_t * const encoder = state->encoder_control; if (!encoder->cfg.alf_type || !alf_enable_luma)//(!cs.sps->getALFEnabledFlag()) || (!alfEnableLuma)) { return; } if (!state->tile->frame->alf_info->ctu_enable_flag[COMPONENT_Y][ctu_rs_addr]) { return; } const unsigned filter_set_idx = state->tile->frame->alf_info->alf_ctb_filter_index[ctu_rs_addr]; unsigned num_aps = state->slice->alf->tile_group_num_aps; unsigned num_available_filt_sets = num_aps + ALF_NUM_FIXED_FILTER_SETS; if (num_available_filt_sets > ALF_NUM_FIXED_FILTER_SETS) { int use_temporal_filt = (filter_set_idx >= ALF_NUM_FIXED_FILTER_SETS) ? 1 : 0; cabac->cur_ctx = &(cabac->ctx.alf_temporal_filt); CABAC_BIN(cabac, use_temporal_filt, "use_latest_filt"); if (use_temporal_filt) { assert(filter_set_idx < num_available_filt_sets); //"temporal non-latest set" if (num_aps > 1) { kvz_cabac_encode_trunc_bin(cabac, filter_set_idx - ALF_NUM_FIXED_FILTER_SETS, num_available_filt_sets - ALF_NUM_FIXED_FILTER_SETS); } } else { assert(filter_set_idx < ALF_NUM_FIXED_FILTER_SETS); //"fixed set larger than temporal" kvz_cabac_encode_trunc_bin(cabac, filter_set_idx, ALF_NUM_FIXED_FILTER_SETS); } } else { assert(filter_set_idx < ALF_NUM_FIXED_FILTER_SETS); //Fixed set numavail < num_fixed kvz_cabac_encode_trunc_bin(cabac, filter_set_idx, ALF_NUM_FIXED_FILTER_SETS); } } static void code_alf_ctu_alternative_ctu(encoder_state_t * const state, cabac_data_t * const cabac, uint32_t ctu_rs_addr, const alf_component_id comp_idx, const alf_aps* aps) { if (comp_idx == COMPONENT_Y) return; int aps_idx = aps ? 0 : state->slice->alf->tile_group_chroma_aps_id; const alf_aps* alf_param_ref = aps ? (aps) : &state->slice->alf->apss[aps_idx]; if (aps || (state->encoder_control->cfg.alf_type && state->slice->alf->tile_group_alf_enabled_flag[comp_idx])) { bool* ctb_alf_flag = state->tile->frame->alf_info->ctu_enable_flag[comp_idx]; if (ctb_alf_flag[ctu_rs_addr]) { const int num_alts = alf_param_ref->num_alternatives_chroma; uint8_t* ctb_alf_alternative = state->tile->frame->alf_info->ctu_alternative[comp_idx]; unsigned num_ones = ctb_alf_alternative[ctu_rs_addr]; assert(ctb_alf_alternative[ctu_rs_addr] < num_alts); for (int i = 0; i < num_ones; ++i) { cabac->cur_ctx = &cabac->ctx.alf_ctb_alternatives[comp_idx - 1]; CABAC_BIN(cabac, 1, "alf_ctb_alternatives"); } if (num_ones < num_alts - 1) { cabac->cur_ctx = &cabac->ctx.alf_ctb_alternatives[comp_idx - 1]; CABAC_BIN(cabac, 0, "alf_ctb_alternatives"); } } } } static void code_alf_ctu_alternatives_component(encoder_state_t * const state, cabac_data_t * const cabac, alf_component_id comp_id, alf_aps* aps) { if (comp_id == COMPONENT_Y) return; uint32_t num_ctus = state->tile->frame->width_in_lcu * state->tile->frame->height_in_lcu; bool* ctb_alf_flag = state->tile->frame->alf_info->ctu_enable_flag[comp_id]; for (int ctu_idx = 0; ctu_idx < num_ctus; ctu_idx++) { if (ctb_alf_flag[ctu_idx]) { code_alf_ctu_alternative_ctu(state, cabac, ctu_idx, comp_id, aps); } } } static void code_alf_ctu_alternatives_channel(encoder_state_t * const state, cabac_data_t * const cabac, channel_type channel, alf_aps* aps) { if (channel == CHANNEL_TYPE_CHROMA) { if (aps->enabled_flag[COMPONENT_Cb]) code_alf_ctu_alternatives_component(state, cabac, COMPONENT_Cb, aps); if (aps->enabled_flag[COMPONENT_Cr]) code_alf_ctu_alternatives_component(state, cabac, COMPONENT_Cr, aps); } } static void code_cc_alf_filter_control_idc(encoder_state_t * const state, cabac_data_t * const cabac, uint8_t idc_val, const alf_component_id comp_id, const int ctu_idx, const uint8_t *filter_control_idc, const int filter_count) { assert(!(idc_val > filter_count)); //Filter index is too large int width_in_lcu = state->tile->frame->width_in_lcu; bool left_avail = ctu_idx % width_in_lcu ? 1 : 0; bool above_avail = ctu_idx / width_in_lcu ? 1 : 0; int ctxt = 0; if (left_avail) { ctxt += (filter_control_idc[ctu_idx - 1]) ? 1 : 0; } if (above_avail) { ctxt += (filter_control_idc[ctu_idx - state->tile->frame->width_in_lcu]) ? 1 : 0; } ctxt += (comp_id == COMPONENT_Cr) ? 3 : 0; //m_BinEncoder.encodeBin((idc_val == 0) ? 0 : 1, Ctx::CcAlfFilterControlFlag(ctxt)); // ON/OFF flag is context coded cabac->cur_ctx = &(cabac->ctx.alf_cc_filter_control_flag[ctxt]); CABAC_BIN(cabac, (idc_val == 0) ? 0 : 1, "cc_alf_filter_control_flag"); if (idc_val > 0) { int val = (idc_val - 1); while (val) { //m_BinEncoder.encodeBinEP(1); CABAC_BIN_EP(cabac, 1, "cc_alf_filter_control_flag"); val--; } if (idc_val < filter_count) { //m_BinEncoder.encodeBinEP(0); CABAC_BIN_EP(cabac, 0, "cc_alf_filter_control_flag"); } } } void kvz_encode_alf_bits(encoder_state_t * const state, const int ctu_idx) { if (state->encoder_control->cfg.alf_type) { alf_info_t *alf_info = state->tile->frame->alf_info; cc_alf_filter_param *cc_filter_param = state->slice->alf->cc_filter_param; bool **ctu_enable_flag = state->tile->frame->alf_info->ctu_enable_flag; for (int comp_idx = 0; comp_idx < MAX_NUM_COMPONENT; comp_idx++) { bool is_luma = comp_idx == COMPONENT_Y ? true : false; //Pitäisi poistaa// /*if (!is_luma) { state->slice->alf->tile_group_alf_enabled_flag[comp_idx] = false; }*/ //---------------// code_alf_ctu_enable_flag(state, &state->cabac, ctu_idx, comp_idx, NULL); if (is_luma) { if (ctu_enable_flag[comp_idx][ctu_idx]) { //int num_aps = state->slice->alf->tile_group_num_aps; //state->slice->alf->tile_group_num_aps = 0; code_alf_ctu_filter_index(state, &state->cabac, ctu_idx, state->slice->alf->tile_group_alf_enabled_flag[COMPONENT_Y]); //state->slice->alf->tile_group_num_aps = num_aps; } } if (!is_luma) { bool* ctb_alf_flag = state->slice->alf->tile_group_alf_enabled_flag[comp_idx] ? ctu_enable_flag[comp_idx] : NULL; if (ctb_alf_flag && ctb_alf_flag[ctu_idx]) { code_alf_ctu_alternative_ctu(state, &state->cabac, ctu_idx, comp_idx, NULL); } } } if (state->encoder_control->cfg.alf_type == KVZ_ALF_FULL) { int num_components = state->encoder_control->chroma_format == KVZ_CSP_400 ? 1 : MAX_NUM_COMPONENT; for (int comp_idx = 1; comp_idx < num_components; comp_idx++) { if (cc_filter_param->cc_alf_filter_enabled[comp_idx - 1]) { const int filter_count = cc_filter_param->cc_alf_filter_count[comp_idx - 1]; code_cc_alf_filter_control_idc(state, &state->cabac, alf_info->cc_alf_filter_control[comp_idx - 1][ctu_idx], comp_idx, ctu_idx, alf_info->cc_alf_filter_control[comp_idx - 1], filter_count); } } } } } static void encode_alf_aps_filter(encoder_state_t * const state, alf_aps* aps, const bool is_chroma, const int alt_idx) { bitstream_t * const stream = &state->stream; const int num_coeff = is_chroma ? 7 : 13; const short* coeff = is_chroma ? aps->chroma_coeff[alt_idx] : aps->luma_coeff; const int16_t* clipp = is_chroma ? aps->chroma_clipp[alt_idx] : aps->luma_clipp; const int num_filters = is_chroma ? 1 : aps->num_luma_filters; // Filter coefficients for (int ind = 0; ind < num_filters; ++ind) { for (int i = 0; i < num_coeff - 1; i++) { WRITE_UE(stream, abs(coeff[ind * MAX_NUM_ALF_LUMA_COEFF + i]), is_chroma ? "alf_chroma_coeff_abs" : "alf_luma_coeff_abs"); //alf_coeff_chroma[i], alf_coeff_luma_delta[i][j] if (abs(coeff[ind* MAX_NUM_ALF_LUMA_COEFF + i]) != 0) { WRITE_U(stream, (coeff[ind * MAX_NUM_ALF_LUMA_COEFF + i] < 0) ? 1 : 0, 1, is_chroma ? "alf_chroma_coeff_sign" : "alf_luma_coeff_sign"); } } } // Clipping values coding if (aps->non_linear_flag[is_chroma]) { for (int ind = 0; ind < num_filters; ++ind) { for (int i = 0; i < num_coeff - 1; i++) { WRITE_U(stream, clipp[ind * MAX_NUM_ALF_LUMA_COEFF + i], 2, is_chroma ? "alf_chroma_clip_idx" : "alf_luma_clip_idx"); } } } } static void encode_alf_aps_flags(encoder_state_t * const state, alf_aps* aps) { bitstream_t * const stream = &state->stream; const bool cc_alf_enabled = state->encoder_control->cfg.alf_type == KVZ_ALF_FULL; WRITE_U(stream, aps->new_filter_flag[CHANNEL_TYPE_LUMA], 1, "alf_luma_new_filter"); if (state->encoder_control->chroma_format != KVZ_CSP_400) { WRITE_U(stream, aps->new_filter_flag[CHANNEL_TYPE_CHROMA], 1, "alf_chroma_new_filter") } if (state->encoder_control->chroma_format != KVZ_CSP_400) { if (cc_alf_enabled) { WRITE_U(stream, aps->cc_alf_aps_param.new_cc_alf_filter[COMPONENT_Cb - 1], 1, "alf_cc_cb_filter_signal_flag"); WRITE_U(stream, aps->cc_alf_aps_param.new_cc_alf_filter[COMPONENT_Cr - 1], 1, "alf_cc_cr_filter_signal_flag"); } else { WRITE_U(stream, 0, 1, "alf_cc_cb_filter_signal_flag"); WRITE_U(stream, 0, 1, "alf_cc_cr_filter_signal_flag"); } } if (aps->new_filter_flag[CHANNEL_TYPE_LUMA]) { WRITE_U(stream, aps->non_linear_flag[CHANNEL_TYPE_LUMA], 1, "alf_luma_clip"); WRITE_UE(stream, aps->num_luma_filters - 1, "alf_luma_num_filters_signalled_minus1"); if (aps->num_luma_filters > 1) { //const int length = ceilLog2(param.numLumaFilters); const int length = kvz_math_ceil_log2(aps->num_luma_filters); for (int i = 0; i < MAX_NUM_ALF_CLASSES; i++) { WRITE_U(stream, aps->filter_coeff_delta_idx[i], length, "alf_luma_coeff_delta_idx"); } } encode_alf_aps_filter(state, aps, false, 0); } if (aps->new_filter_flag[CHANNEL_TYPE_CHROMA]) { WRITE_U(stream, aps->non_linear_flag[CHANNEL_TYPE_CHROMA], 1, "alf_nonlinear_enable_flag_chroma"); if (MAX_NUM_ALF_ALTERNATIVES_CHROMA > 1) { //WRITE_UVLC(param.numAlternativesChroma - 1, "alf_chroma_num_alts_minus1"); WRITE_UE(stream, aps->num_alternatives_chroma - 1, "alf_chroma_num_alts_minus1"); } for (int alt_idx = 0; alt_idx < aps->num_alternatives_chroma; ++alt_idx) { encode_alf_aps_filter(state, aps, true, alt_idx); } } if (cc_alf_enabled) { for (int cc_idx = 0; cc_idx < 2; cc_idx++) { if (aps->cc_alf_aps_param.new_cc_alf_filter[cc_idx]) { const int filter_count = aps->cc_alf_aps_param.cc_alf_filter_count[cc_idx]; assert(filter_count <= MAX_NUM_CC_ALF_FILTERS); // "CC ALF Filter count is too large" assert(filter_count > 0); // "CC ALF Filter count is too small" if (MAX_NUM_CC_ALF_FILTERS > 1) { WRITE_UE(stream, filter_count - 1, cc_idx == 0 ? "alf_cc_cb_filters_signalled_minus1" : "alf_cc_cr_filters_signalled_minus1"); } for (int filter_idx = 0; filter_idx < filter_count; filter_idx++) { int num_coeff = MAX_NUM_CC_ALF_CHROMA_COEFF; //CC_ALF_FILTER const short *coeff = aps->cc_alf_aps_param.cc_alf_coeff[cc_idx][filter_idx]; // Filter coefficients for (int i = 0; i < num_coeff - 1; i++) { if (coeff[i] == 0) { WRITE_U(stream, 0, 3, cc_idx == 0 ? "alf_cc_cb_mapped_coeff_abs" : "alf_cc_cr_mapped_coeff_abs"); } else { WRITE_U(stream, 1 + kvz_math_floor_log2(abs(coeff[i])), 3, cc_idx == 0 ? "alf_cc_cb_mapped_coeff_abs" : "alf_cc_cr_mapped_coeff_abs"); WRITE_U(stream, coeff[i] < 0 ? 1 : 0, 1, cc_idx == 0 ? "alf_cc_cb_coeff_sign" : "alf_cc_cr_coeff_sign"); } } } } } } } // ToDo: Fill in LMCS APS static void encode_lmcs_aps(encoder_state_t* const state, lmcs_aps* aps) { bitstream_t* const stream = &state->stream; //SliceReshapeInfo param = pcAPS->getReshaperAPSInfo(); WRITE_UE(stream, 0/*param.reshaperModelMinBinIdx*/, "lmcs_min_bin_idx"); WRITE_UE(stream, 16 - 1/*16 - 1 - param.reshaperModelMaxBinIdx*/, "lmcs_delta_max_bin_idx"); WRITE_UE(stream, 7/*param.maxNbitsNeededDeltaCW - 1*/, "lmcs_delta_cw_prec_minus1"); /* for (int i = param.reshaperModelMinBinIdx; i <= param.reshaperModelMaxBinIdx; i++) { int deltaCW = param.reshaperModelBinCWDelta[i]; int signCW = (deltaCW < 0) ? 1 : 0; int absCW = (deltaCW < 0) ? (-deltaCW) : deltaCW; WRITE_CODE(absCW, param.maxNbitsNeededDeltaCW, "lmcs_delta_abs_cw[ i ]"); if (absCW > 0) { WRITE_FLAG(signCW, "lmcs_delta_sign_cw_flag[ i ]"); } } int deltaCRS = pcAPS->chromaPresentFlag ? param.chrResScalingOffset : 0; int signCRS = (deltaCRS < 0) ? 1 : 0; int absCRS = (deltaCRS < 0) ? (-deltaCRS) : deltaCRS; if (pcAPS->chromaPresentFlag) { WRITE_CODE(absCRS, 3, "lmcs_delta_abs_crs"); } if (absCRS > 0) { WRITE_FLAG(signCRS, "lmcs_delta_sign_crs_flag"); } */ } static void encoder_state_write_adaptation_parameter_set(encoder_state_t * const state, alf_aps *aps) { #ifdef KVZ_DEBUG printf("=========== Adaptation Parameter Set ===========\n"); #endif bitstream_t * const stream = &state->stream; WRITE_U(stream, (int)aps->aps_type, 3, "aps_params_type"); WRITE_U(stream, aps->aps_id, 5, "adaptation_parameter_set_id"); WRITE_U(stream, state->encoder_control->chroma_format != KVZ_CSP_400, 1, "aps_chroma_present_flag"); if (aps->aps_type == T_ALF_APS) { encode_alf_aps_flags(state, aps); } else if (aps->aps_type == T_LMCS_APS) { //encode_lmcs_aps(state); } /*else if (aps->aps_type == T_SCALING_LIST_APS) { codeScalingListAps(pcAPS); }*/ WRITE_U(stream, 0, 1, "aps_extension_flag"); //Implementation when this flag is equal to 1 should be added when it is needed. Currently in the spec we don't have case when this flag is equal to 1 kvz_bitstream_add_rbsp_trailing_bits(stream); } static void encode_alf_aps(encoder_state_t * const state) { const encoder_control_t * const encoder = state->encoder_control; bitstream_t * const stream = &state->stream; if (encoder->cfg.alf_type && (state->slice->alf->tile_group_alf_enabled_flag[COMPONENT_Y] || state->slice->alf->tile_group_cc_alf_cb_enabled_flag || state->slice->alf->tile_group_cc_alf_cr_enabled_flag)) { param_set_map *aps_map = state->tile->frame->alf_param_set_map; for (int aps_id = 0; aps_id < ALF_CTB_MAX_NUM_APS; aps_id++) { alf_aps aps = aps_map[aps_id + T_ALF_APS + NUM_APS_TYPE_LEN].parameter_set; bool write_aps = aps_map[aps_id + T_ALF_APS + NUM_APS_TYPE_LEN].b_changed; /*if (!write_aps && state->slice->alf->apss && state->slice->alf->apss[aps_id].aps_id >= 0 && state->slice->alf->apss[aps_id].aps_id < 8) { write_aps = true; aps = state->slice->alf->apss[aps_id]; // use aps from slice header //*apsMap->allocatePS(apsId) = *aps; //allocate and cpy copy_aps_to_map(aps_map, &aps, aps_id + T_ALF_APS + NUM_APS_TYPE_LEN); //m_pcALF->setApsIdStart(apsId); g_aps_id_start = aps_id; }*/ if (write_aps) { kvz_nal_write(stream, NAL_UNIT_PREFIX_APS, 0, state->frame->first_nal); state->frame->first_nal = false; encoder_state_write_adaptation_parameter_set(state, &aps); aps_map[aps_id + T_ALF_APS].b_changed = false; } } } } void kvz_encode_alf_adaptive_parameter_set(encoder_state_t * const state) { //send LMCS APS when LMCSModel is updated. It can be updated even current slice does not enable reshaper. //For example, in RA, update is on intra slice, but intra slice may not use reshaper //encode_alf_aps_lmcs(state); // only 1 SCALING LIST data for 1 picture //encode_alf_aps_scaling_list(state); encode_alf_aps(state); } //-------------------------------------------------------------------------- //-------------------------CC ALF encoding functions------------------------ static void filter_blk_cc_alf(encoder_state_t * const state, kvz_pixel *dst_buf, const kvz_pixel *rec_src, const int rec_luma_stride, const alf_component_id comp_id, const int16_t *filter_coeff, const clp_rngs clp_rngs, int vb_ctu_height, int vb_pos, const int x_pos, const int y_pos, const int blk_width, const int blk_height) { assert(!(1 << kvz_math_floor_log2(vb_ctu_height) != vb_ctu_height)); //Not a power of 2 assert(comp_id != COMPONENT_Y); //Must be chroma enum kvz_chroma_format chroma_format = state->encoder_control->chroma_format; uint8_t scale_y = (comp_id == COMPONENT_Y || chroma_format != KVZ_CSP_420) ? 0 : 1; uint8_t scale_x = (comp_id == COMPONENT_Y || chroma_format == KVZ_CSP_444) ? 0 : 1; const int cls_size_y = 4; const int cls_size_x = 4; const int start_height = y_pos; const int end_height = y_pos + blk_height; const int start_width = x_pos; const int end_width = x_pos + blk_width; const int luma_start_height = start_height << scale_y; const int luma_start_width = start_width << scale_x; assert(!(start_height % cls_size_y)); //Wrong start_height in filtering assert(!(start_width % cls_size_x)); //Wrong start_width in filtering assert(!((end_height - start_height) % cls_size_y)); //Wrong end_height in filtering assert(!((end_width - start_width) % cls_size_x)); //Wrong end_width in filtering const kvz_pixel* src_buf = rec_src; const kvz_pixel* luma_ptr = src_buf + luma_start_height * rec_luma_stride + luma_start_width; const int chroma_stride = rec_luma_stride >> scale_x; kvz_pixel* chroma_ptr = dst_buf + start_height * chroma_stride + start_width; for (int i = 0; i < end_height - start_height; i += cls_size_y) { for (int j = 0; j < end_width - start_width; j += cls_size_x) { for (int ii = 0; ii < cls_size_y; ii++) { int row = ii; int col = j; kvz_pixel *src_self = chroma_ptr + col + row * chroma_stride; int offset1 = rec_luma_stride; int offset2 = -rec_luma_stride; int offset3 = 2 * rec_luma_stride; row <<= scale_y; col <<= scale_x; const kvz_pixel *src_cross = luma_ptr + col + row * rec_luma_stride; int pos = ((start_height + i + ii) << scale_y) & (vb_ctu_height - 1); if (scale_y == 0 && (pos == vb_pos || pos == vb_pos + 1)) { continue; } if (pos == (vb_pos - 2) || pos == (vb_pos + 1)) { offset3 = offset1; } else if (pos == (vb_pos - 1) || pos == vb_pos) { offset1 = 0; offset2 = 0; offset3 = 0; } for (int jj = 0; jj < cls_size_x; jj++) { const int jj2 = (jj << scale_x); const int offset0 = 0; int sum = 0; const kvz_pixel curr_src_cross = src_cross[offset0 + jj2]; sum += filter_coeff[0] * (src_cross[offset2 + jj2] - curr_src_cross); sum += filter_coeff[1] * (src_cross[offset0 + jj2 - 1] - curr_src_cross); sum += filter_coeff[2] * (src_cross[offset0 + jj2 + 1] - curr_src_cross); sum += filter_coeff[3] * (src_cross[offset1 + jj2 - 1] - curr_src_cross); sum += filter_coeff[4] * (src_cross[offset1 + jj2] - curr_src_cross); sum += filter_coeff[5] * (src_cross[offset1 + jj2 + 1] - curr_src_cross); sum += filter_coeff[6] * (src_cross[offset3 + jj2] - curr_src_cross); sum = (sum + ((1 << 7/*m_scaleBits*/) >> 1)) >> 7/*m_scaleBits*/; const int offset = 1 << clp_rngs.comp[comp_id].bd >> 1; sum = kvz_fast_clip_32bit_to_pixel(sum + offset) - offset; sum += src_self[jj]; src_self[jj] = kvz_fast_clip_32bit_to_pixel(sum); } } } chroma_ptr += chroma_stride * cls_size_y; luma_ptr += rec_luma_stride * cls_size_y << scale_y; } } static void apply_cc_alf_filter(encoder_state_t * const state, alf_component_id comp_id, kvz_pixel *dst_buf, const kvz_pixel *rec_yuv_ext, const int luma_stride, uint8_t *filter_control, const short filter_set[MAX_NUM_CC_ALF_FILTERS][MAX_NUM_CC_ALF_CHROMA_COEFF], const int selected_filter_idx, array_variables *arr_vars) { enum kvz_chroma_format chroma_format = state->encoder_control->chroma_format; uint8_t component_scale_y = (comp_id == COMPONENT_Y || chroma_format != KVZ_CSP_420) ? 0 : 1; uint8_t component_scale_x = (comp_id == COMPONENT_Y || chroma_format == KVZ_CSP_444) ? 0 : 1; const int pic_height = state->tile->frame->height; const int pic_width = state->tile->frame->width; const int max_ctu_height_log2 = kvz_math_floor_log2(LCU_WIDTH); const int max_ctu_width_log2 = kvz_math_floor_log2(LCU_WIDTH); const int width_in_ctus = state->tile->frame->width_in_lcu; const int alf_vb_luma_ctu_height = LCU_WIDTH; const int alf_vb_luma_pos = LCU_WIDTH - ALF_VB_POS_ABOVE_CTUROW_LUMA; int ctu_idx = 0; for (int y_pos = 0; y_pos < pic_height; y_pos += LCU_WIDTH) { for (int x_pos = 0; x_pos < pic_width; x_pos += LCU_WIDTH) { int filter_idx = (filter_control == NULL) ? selected_filter_idx : filter_control[(y_pos >> max_ctu_height_log2) * width_in_ctus + (x_pos >> max_ctu_width_log2)]; bool skip_filtering = (filter_control != NULL && filter_idx == 0) ? true : false; if (!skip_filtering) { if (filter_control != NULL) { filter_idx--; } const int16_t *filter_coeff = filter_set[filter_idx]; const int width = (x_pos + LCU_WIDTH > pic_width) ? (pic_width - x_pos) : LCU_WIDTH; const int height = (y_pos + LCU_WIDTH > pic_height) ? (pic_height - y_pos) : LCU_WIDTH; { filter_blk_cc_alf(state, dst_buf, rec_yuv_ext, luma_stride, comp_id, filter_coeff, arr_vars->clp_rngs, alf_vb_luma_ctu_height, alf_vb_luma_pos, x_pos >> component_scale_x, y_pos >> component_scale_y, width >> component_scale_x, height >> component_scale_y); } } ctu_idx++; } } } static void setup_cc_alf_aps(encoder_state_t * const state, const int *cc_reuse_aps_id) { cc_alf_filter_param *cc_filter_param = state->slice->alf->cc_filter_param; if (cc_filter_param->cc_alf_filter_enabled[COMPONENT_Cb - 1]) { int cc_alf_cb_aps_id = state->slice->alf->tile_group_cc_alf_cb_aps_id; alf_aps *aps = &state->tile->frame->alf_param_set_map[cc_alf_cb_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set; if (aps->aps_id >= 0 && aps->aps_id < ALF_CTB_MAX_NUM_APS) { //aps = m_apsMap->allocatePS((cc_alf_cb_aps_id << NUM_APS_TYPE_LEN) + ALF_APS); aps->temporal_id = 0; // cs.slice->getTLayer() } aps->cc_alf_aps_param.cc_alf_filter_enabled[COMPONENT_Cb - 1] = 1; aps->cc_alf_aps_param.cc_alf_filter_count[COMPONENT_Cb - 1] = cc_filter_param->cc_alf_filter_count[COMPONENT_Cb - 1]; for (int filter_idx = 0; filter_idx < MAX_NUM_CC_ALF_FILTERS; filter_idx++) { aps->cc_alf_aps_param.cc_alf_filter_idx_enabled[COMPONENT_Cb - 1][filter_idx] = cc_filter_param->cc_alf_filter_idx_enabled[COMPONENT_Cb - 1][filter_idx]; memcpy(aps->cc_alf_aps_param.cc_alf_coeff[COMPONENT_Cb - 1][filter_idx], cc_filter_param->cc_alf_coeff[COMPONENT_Cb - 1][filter_idx], sizeof(short) * MAX_NUM_CC_ALF_CHROMA_COEFF); } aps->aps_id = cc_alf_cb_aps_id; aps->aps_type = T_ALF_APS; if (cc_reuse_aps_id[COMPONENT_Cb - 1] < 0) { aps->cc_alf_aps_param.new_cc_alf_filter[COMPONENT_Cb - 1] = 1; state->tile->frame->alf_param_set_map[cc_alf_cb_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].b_changed = true; aps->temporal_id = 0; // cs.slice->getTLayer() } state->slice->alf->tile_group_cc_alf_cb_enabled_flag = true; } else { state->slice->alf->tile_group_cc_alf_cb_enabled_flag = false; } if (cc_filter_param->cc_alf_filter_enabled[COMPONENT_Cr - 1]) { int cc_alf_cr_aps_id = state->slice->alf->tile_group_cc_alf_cr_aps_id; alf_aps *aps = &state->tile->frame->alf_param_set_map[cc_alf_cr_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set; if (aps->aps_id >= 0 && aps->aps_id < ALF_CTB_MAX_NUM_APS) { //aps = m_apsMap->allocatePS((cc_alf_cb_aps_id << NUM_APS_TYPE_LEN) + ALF_APS); aps->temporal_id = 0; // cs.slice->getTLayer() } aps->cc_alf_aps_param.cc_alf_filter_enabled[COMPONENT_Cr - 1] = 1; aps->cc_alf_aps_param.cc_alf_filter_count[COMPONENT_Cr - 1] = cc_filter_param->cc_alf_filter_count[COMPONENT_Cr - 1]; for (int filter_idx = 0; filter_idx < MAX_NUM_CC_ALF_FILTERS; filter_idx++) { aps->cc_alf_aps_param.cc_alf_filter_idx_enabled[COMPONENT_Cr - 1][filter_idx] = cc_filter_param->cc_alf_filter_idx_enabled[COMPONENT_Cr - 1][filter_idx]; memcpy(aps->cc_alf_aps_param.cc_alf_coeff[COMPONENT_Cr - 1][filter_idx], cc_filter_param->cc_alf_coeff[COMPONENT_Cr - 1][filter_idx], sizeof(short) * MAX_NUM_CC_ALF_CHROMA_COEFF); } aps->aps_id = cc_alf_cr_aps_id; aps->aps_type = T_ALF_APS; if (cc_reuse_aps_id[COMPONENT_Cr - 1] < 0) { aps->cc_alf_aps_param.new_cc_alf_filter[COMPONENT_Cr - 1] = 1; state->tile->frame->alf_param_set_map[cc_alf_cr_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].b_changed = true; aps->temporal_id = 0; // cs.slice->getTLayer() } state->slice->alf->tile_group_cc_alf_cr_enabled_flag = true; } else { state->slice->alf->tile_group_cc_alf_cr_enabled_flag = false; } } static void round_filt_coeff_cc_alf(int16_t *filter_coeff_quant, double *filter_coeff, const int num_coeff, const int factor) { for (int i = 0; i < num_coeff; i++) { int sign = filter_coeff[i] > 0 ? 1 : -1; double best_err = 128.0*128.0; int best_index = 0; for (int k = 0; k < CCALF_CANDS_COEFF_NR; k++) { double err = (filter_coeff[i] * sign * factor - cc_alf_small_tab[k]); err = err * err; if (err < best_err) { best_err = err; best_index = k; } } filter_coeff_quant[i] = cc_alf_small_tab[best_index] * sign; } } static int get_coeff_rate_cc_alf(short chroma_coeff[MAX_NUM_CC_ALF_FILTERS][MAX_NUM_CC_ALF_CHROMA_COEFF], bool filter_enabled[MAX_NUM_CC_ALF_FILTERS], uint8_t filter_count, alf_component_id comp_id) { int bits = 0; if (filter_count > 0) { bits += length_uvlc(filter_count - 1); int signaled_filter_count = 0; for (int filterIdx = 0; filterIdx < MAX_NUM_CC_ALF_FILTERS; filterIdx++) { if (filter_enabled[filterIdx]) { // Filter coefficients for (int i = 0; i < CC_ALF_NUM_COEFF - 1; i++) { bits += CCALF_BITS_PER_COEFF_LEVEL + (chroma_coeff[filterIdx][i] == 0 ? 0 : 1); } signaled_filter_count++; } } assert(signaled_filter_count == filter_count); //Number of filter signaled not same as indicated } return bits; } static void derive_cc_alf_filter_coeff(alf_covariance *alf_covariance_frame_cc_alf, short filter_coeff[MAX_NUM_CC_ALF_FILTERS][MAX_NUM_CC_ALF_CHROMA_COEFF], const uint8_t filter_idx) { int forward_tab[CCALF_CANDS_COEFF_NR * 2 - 1] = { 0 }; for (int i = 0; i < CCALF_CANDS_COEFF_NR; i++) { forward_tab[CCALF_CANDS_COEFF_NR - 1 + i] = cc_alf_small_tab[i]; forward_tab[CCALF_CANDS_COEFF_NR - 1 - i] = (-1) * cc_alf_small_tab[i]; } double filter_coeff_dbl[MAX_NUM_CC_ALF_CHROMA_COEFF]; int16_t filter_coeff_int[MAX_NUM_CC_ALF_CHROMA_COEFF]; memset(filter_coeff_int, 0, sizeof(filter_coeff_int)); double k_e[MAX_NUM_ALF_LUMA_COEFF][MAX_NUM_ALF_LUMA_COEFF]; double ky[MAX_NUM_ALF_LUMA_COEFF]; const int size = CC_ALF_NUM_COEFF - 1; for (int k = 0; k < size; k++) { ky[k] = alf_covariance_frame_cc_alf[filter_idx].y[0][k]; for (int l = 0; l < size; l++) { k_e[k][l] = alf_covariance_frame_cc_alf[filter_idx].ee[0][0][k][l]; } } //m_alfCovarianceFrameCcAlf[compID - 1][0][filterIdx].gnsSolveByChol(kE, ky, filterCoeffDbl, size); gns_solve_by_chol(k_e, ky, filter_coeff_dbl, size); round_filt_coeff_cc_alf(filter_coeff_int, filter_coeff_dbl, size, (1 << 7 /*m_scaleBits*/)); for (int k = 0; k < size; k++) { assert(!(filter_coeff_int[k] < -(1 << CCALF_DYNAMIC_RANGE))); // this is not possible: filter_coeff_int[k] < -(1 << CCALF_DYNAMIC_RANGE) assert(!(filter_coeff_int[k] > (1 << CCALF_DYNAMIC_RANGE))); // this is not possible: filter_coeff_int[k] > (1 << CCALF_DYNAMIC_RANGE) } // Refine quanitzation int modified = 1; double err_ref = calc_error_for_cc_alf_coeffs(&alf_covariance_frame_cc_alf[filter_idx], filter_coeff_int, size, (7/*m_scaleBits*/ + 1)); while (modified) { modified = 0; for (int i = 1; i > -2; i -= 2) { int delta = i; double err_min = MAX_DOUBLE; int idx_min = -1; int min_index = -1; for (int k = 0; k < size; k++) { int org_idx = -1; for (int i = 0; i < CCALF_CANDS_COEFF_NR * 2 - 1; i++) { if (forward_tab[i] == filter_coeff_int[k]) { org_idx = i; break; } } assert(!(org_idx < 0)); //this is wrong, does not find coeff from forward_tab if ((org_idx - delta < 0) || (org_idx - delta >= CCALF_CANDS_COEFF_NR * 2 - 1)) continue; filter_coeff_int[k] = forward_tab[org_idx - delta]; double error = calc_error_for_cc_alf_coeffs(&alf_covariance_frame_cc_alf[filter_idx], filter_coeff_int, size, (7/*m_scaleBits*/ + 1)); if (error < err_min) { err_min = error; idx_min = k; min_index = org_idx; } filter_coeff_int[k] = forward_tab[org_idx]; } if (err_min < err_ref) { min_index -= delta; assert(!(min_index < 0));// this is wrong, index - delta < 0 assert(!(min_index >= CCALF_CANDS_COEFF_NR * 2 - 1)); // this is wrong, index - delta >= CCALF_CANDS_COEFF_NR * 2 - 1 filter_coeff_int[idx_min] = forward_tab[min_index]; modified++; err_ref = err_min; } } } for (int k = 0; k < (size + 1); k++) { assert(!((filter_coeff_int[k] < -(1 << CCALF_DYNAMIC_RANGE)) || (filter_coeff_int[k] > (1 << CCALF_DYNAMIC_RANGE)))); //Exceeded valid range for CC ALF coefficient filter_coeff[filter_idx][k] = filter_coeff_int[k]; } } static void determine_control_idc_values(encoder_state_t *const state, const alf_component_id comp_id, const int ctu_width_c, const int ctu_height_c, const int pic_width_c, const int pic_height_c, double **unfiltered_distortion, uint64_t *training_distortion[MAX_NUM_CC_ALF_FILTERS], bool reuse_temporal_filter_coeff, uint8_t *training_cov_control, uint8_t *filter_control, uint64_t *cur_total_distortion, double *cur_total_rate, bool filter_enabled[MAX_NUM_CC_ALF_FILTERS], uint8_t map_filter_idx_to_filter_idc[MAX_NUM_CC_ALF_FILTERS + 1], uint8_t *cc_alf_filter_count) { bool cur_filter_enabled[MAX_NUM_CC_ALF_FILTERS]; //std::fill_n(cur_filter_enabled, MAX_NUM_CC_ALF_FILTERS, false); memset(cur_filter_enabled, false, sizeof(cur_filter_enabled)); #if MAX_NUM_CC_ALF_FILTERS>1 filter_idx_count filter_idx_count[MAX_NUM_CC_ALF_FILTERS]; for (int i = 0; i < MAX_NUM_CC_ALF_FILTERS; i++) { filter_idx_count[i].count = 0; filter_idx_count[i].filter_idx = i; } double prev_rate = (*cur_total_rate); #endif alf_info_t *alf_info = state->tile->frame->alf_info; cabac_data_t *cabac_estimator = &alf_info->cabac_estimator; cabac_data_t ctx_initial; cabac_data_t ctx_best; cabac_data_t ctx_start; memcpy(&ctx_initial, cabac_estimator, sizeof(ctx_initial)); memcpy(&ctx_best, cabac_estimator, sizeof(ctx_best)); cabac_estimator->only_count = 1; ctx_initial.only_count = 1; ctx_best.only_count = 1; ctx_start.only_count = 1; //enum kvz_chroma_format chroma_format = state->encoder_control->chroma_format; double lambda = state->frame->lambda; bool limit_cc_alf = state->encoder_control->cfg.qp >= 37; int ctu_idx = 0; for (int y_ctu = 0; y_ctu < pic_height_c; y_ctu += ctu_height_c) { for (int x_ctu = 0; x_ctu < pic_width_c; x_ctu += ctu_width_c) { uint64_t ssd; double rate; double cost; uint64_t best_ssd = (MAX_INT64 >> 1); double best_rate = MAX_DOUBLE; double best_cost = MAX_DOUBLE; uint8_t best_filter_idc = 0; uint8_t best_filter_idx = 0; //const uint32_t threshold_s = MIN(pic_height_c - y_ctu, ctu_height_c) << (chroma_format != KVZ_CSP_420 ? 0 : 1); //const uint32_t number_of_chroma_samples = MIN(pic_height_c - y_ctu, ctu_height_c) * MIN(pic_height_c - x_ctu, ctu_width_c); //const uint32_t threshold_c = (number_of_chroma_samples >> 2); memcpy(cabac_estimator, &ctx_best, sizeof(*cabac_estimator)); memcpy(&ctx_start, cabac_estimator, sizeof(ctx_start)); for (int filter_idx = 0; filter_idx <= MAX_NUM_CC_ALF_FILTERS; filter_idx++) { uint8_t filter_idc = map_filter_idx_to_filter_idc[filter_idx]; if (filter_idx < MAX_NUM_CC_ALF_FILTERS && !filter_enabled[filter_idx]) { continue; } if (filter_idx == MAX_NUM_CC_ALF_FILTERS) { ssd = (uint64_t)unfiltered_distortion[comp_id][ctu_idx]; // restore saved distortion computation } else { ssd = training_distortion[filter_idx][ctu_idx]; } memcpy(cabac_estimator, &ctx_start, sizeof(*cabac_estimator)); alf_cabac_reset_bits(cabac_estimator); //const Position lumaPos = Position({ xCtu << getComponentScaleX(comp_id, cs.pcv->chrFormat), // yCtu << getComponentScaleY(comp_id, cs.pcv->chrFormat) }); code_cc_alf_filter_control_idc(state, cabac_estimator, filter_idc, comp_id, ctu_idx, filter_control, *cc_alf_filter_count); //rate = FRAC_BITS_SCALE * m_CABACEstimator->getEstFracBits(); rate = (23 - cabac_estimator->bits_left) + (cabac_estimator->num_buffered_bytes << 3); cost = rate * lambda + ssd; bool limitation_exceeded = false; if (limit_cc_alf && filter_idx < MAX_NUM_CC_ALF_FILTERS) { assert(false); // should have returned from cc alf with limit_cc_alf == true //limitation_exceeded = limitation_exceeded || (luma_swing_greater_than_threshold_count[ctu_idx] >= threshold_s); //limitation_exceeded = limitation_exceeded || (chroma_sample_count_near_mid_point[ctu_idx] >= threshold_c); } if (cost < best_cost && !limitation_exceeded) { best_cost = cost; best_rate = rate; best_ssd = ssd; best_filter_idc = filter_idc; best_filter_idx = filter_idx; //ctx_best = SubCtx(Ctx::CcAlfFilterControlFlag, m_CABACEstimator->getCtx()); memcpy(&ctx_best, cabac_estimator, sizeof(ctx_best)); training_cov_control[ctu_idx] = (filter_idx == MAX_NUM_CC_ALF_FILTERS) ? 0 : (filter_idx + 1); filter_control[ctu_idx] = (filter_idx == MAX_NUM_CC_ALF_FILTERS) ? 0 : (filter_idx + 1); } } if (best_filter_idc != 0) { cur_filter_enabled[best_filter_idx] = true; #if MAX_NUM_CC_ALF_FILTERS>1 filter_idx_count[best_filter_idx].count++; #endif } (*cur_total_rate) += best_rate; (*cur_total_distortion) += best_ssd; ctu_idx++; } } #if MAX_NUM_CC_ALF_FILTERS>1 if (!reuse_temporal_filter_coeff) { memcpy(cur_filter_enabled, filter_enabled, sizeof(cur_filter_enabled)); qsort(filter_idx_count, MAX_NUM_CC_ALF_FILTERS, sizeof(*filter_idx_count), comparator); int filter_idc = 1; (*cc_alf_filter_count) = 0; for (int i = 0; i < MAX_NUM_CC_ALF_FILTERS; i++) { const int filter_idx = filter_idx_count[i].filter_idx; if (filter_enabled[filter_idx]) { map_filter_idx_to_filter_idc[filter_idx] = filter_idc; filter_idc++; (*cc_alf_filter_count)++; } } (*cur_total_rate) = prev_rate; //m_CABACEstimator->getCtx() = ctx_initial; memcpy(cabac_estimator, &ctx_initial, sizeof(*cabac_estimator)); //m_CABACEstimator->resetBits(); alf_cabac_reset_bits(cabac_estimator); int ctu_idx = 0; for (int y = 0; y < pic_height_c; y += ctu_height_c) { for (int x = 0; x < pic_width_c; x += ctu_width_c) { const int filter_idx_plus1 = filter_control[ctu_idx]; code_cc_alf_filter_control_idc(state, cabac_estimator, (filter_idx_plus1 == 0 ? 0 : map_filter_idx_to_filter_idc[filter_idx_plus1 - 1]), comp_id, ctu_idx, filter_control, *cc_alf_filter_count); ctu_idx++; } } (*cur_total_rate) += (23 - cabac_estimator->bits_left) + (cabac_estimator->num_buffered_bytes << 3); } #endif // restore for next iteration memcpy(cabac_estimator, &ctx_initial, sizeof(*cabac_estimator)); } static void get_available_cc_alf_aps_ids(encoder_state_t *const state, alf_component_id compID, int *aps_ids_size, int *aps_ids) { for (int i = 0; i < ALF_CTB_MAX_NUM_APS; i++) { param_set_map* param_set = &state->tile->frame->alf_param_set_map[i + NUM_APS_TYPE_LEN + T_ALF_APS]; if (param_set->b_changed && (param_set->parameter_set.aps_id >= 0 || param_set->parameter_set.aps_id < ALF_CTB_MAX_NUM_APS)) { copy_aps(&state->slice->alf->apss[i], ¶m_set->parameter_set, true); } } int aps_id_checked = 0, cur_aps_id = state->tile->frame->alf_info->aps_id_start; if (cur_aps_id < ALF_CTB_MAX_NUM_APS) { while (aps_id_checked < ALF_CTB_MAX_NUM_APS && !state->frame->is_irap && (*aps_ids_size) < ALF_CTB_MAX_NUM_APS /*&& !cs.slice->getPendingRasInit()*/) { alf_aps cur_aps = state->slice->alf->apss[cur_aps_id]; bool aps_found = (0 <= cur_aps.aps_id && cur_aps.aps_id < ALF_CTB_MAX_NUM_APS); if (aps_found && cur_aps.temporal_id <= state->slice->id && cur_aps.cc_alf_aps_param.new_cc_alf_filter[compID - 1]) { aps_ids[(*aps_ids_size)] = cur_aps_id; (*aps_ids_size) += 1; } aps_id_checked++; cur_aps_id = (cur_aps_id + 1) % ALF_CTB_MAX_NUM_APS; } } } static void get_frame_stats_cc_alf(alf_covariance* alf_covariance_cc_alf, alf_covariance* alf_covariance_frame_cc_alf, int filter_idc, const int num_ctus_in_frame, uint8_t *training_cov_control) { const int filter_idx = filter_idc - 1; // init Frame stats buffers reset_alf_covariance(&alf_covariance_frame_cc_alf[filter_idx], -1); for (int ctu_rs_addr = 0; ctu_rs_addr < num_ctus_in_frame; ctu_rs_addr++) { if (training_cov_control[ctu_rs_addr] == filter_idc) { add_alf_cov(&alf_covariance_frame_cc_alf[filter_idx], &alf_covariance_cc_alf[ctu_rs_addr]); } } } static void derive_cc_alf_filter(encoder_state_t * const state, alf_component_id comp_id, const kvz_picture *org_yuv, const kvz_picture *rec_dst_yuv, int *cc_reuse_aps_id) { cc_alf_filter_param *cc_filter_param = state->slice->alf->cc_filter_param; if (!state->slice->alf->tile_group_alf_enabled_flag[COMPONENT_Y]) { cc_filter_param->cc_alf_filter_enabled[comp_id - 1] = false; return; } bool limit_cc_alf = state->encoder_control->cfg.qp >= 37; // m_encCfg->getCCALFQpThreshold(); if (limit_cc_alf) // && state->slice. cs.slice->getSliceQp() <= m_encCfg->getBaseQP() + 1) { cc_filter_param->cc_alf_filter_enabled[comp_id - 1] = false; return; } alf_info_t *alf_info = state->tile->frame->alf_info; alf_covariance *alf_covariance_cc_alf = state->tile->frame->alf_info->alf_covariance_cc_alf[comp_id - 1]; alf_covariance *alf_covariance_frame_cc_alf = state->tile->frame->alf_info->alf_covariance_frame_cc_alf[comp_id - 1]; uint8_t* training_cov_control = alf_info->training_cov_control; uint8_t* filter_control = alf_info->filter_control; uint8_t* best_filter_control = alf_info->best_filter_control; enum kvz_chroma_format chroma_fmt = state->encoder_control->chroma_format; uint8_t best_map_filter_idx_to_filter_idc[MAX_NUM_CC_ALF_FILTERS + 1]; bool scale_x = (comp_id == COMPONENT_Y || chroma_fmt == KVZ_CSP_444) ? 0 : 1; bool scale_y = (comp_id == COMPONENT_Y || chroma_fmt != KVZ_CSP_420) ? 0 : 1; const int ctu_width_c = LCU_WIDTH >> scale_x; const int ctu_height_c = LCU_WIDTH >> scale_y; const int pic_width_c = state->tile->frame->width >> scale_x; const int pic_height_c = state->tile->frame->height >> scale_y; //const int pic_stride_c = rec_dst_yuv->stride >> scale_x; //const int8_t bit_depth = state->encoder_control->bitdepth; const int max_training_iter_count = 15; int max_ctu_height_log2 = kvz_math_floor_log2(LCU_WIDTH); //int max_ctu_height_log2_chrma = kvz_math_floor_log2(LCU_WIDTH) >> scale_y; int max_ctu_width_log2 = kvz_math_floor_log2(LCU_WIDTH); //int max_ctu_width_log2_chrma = kvz_math_floor_log2(LCU_WIDTH) >> scale_x; int32_t ctus_in_width = state->tile->frame->width_in_lcu; const uint32_t num_ctus_in_pic = state->tile->frame->width_in_lcu * state->tile->frame->height_in_lcu; short best_filter_coeff_set[MAX_NUM_CC_ALF_FILTERS][MAX_NUM_CC_ALF_CHROMA_COEFF]; bool best_filter_idx_enabled[MAX_NUM_CC_ALF_FILTERS]; uint8_t best_filter_count = 0; double lambda = state->frame->lambda; cabac_data_t *cabac_estimator = &alf_info->cabac_estimator; cabac_data_t ctx_start_cc_alf_filter_control_flag; //uint64_t* luma_swing_greater_than_threshold_count; //uint64_t* chroma_sample_count_near_mid_point; /*if (limit_cc_alf) { luma_swing_greater_than_threshold_count = malloc(num_ctus_in_pic * sizeof(*luma_swing_greater_than_threshold_count)); count_luma_swing_greater_than_threshold(rec_dst_yuv->y, rec_dst_yuv->stride, rec_dst_yuv->height, rec_dst_yuv->width, max_ctu_width_log2, max_ctu_height_log2, luma_swing_greater_than_threshold_count, ctus_in_width, bit_depth); } if (limit_cc_alf) { chroma_sample_count_near_mid_point = malloc(num_ctus_in_pic * sizeof(*chroma_sample_count_near_mid_point)); if (comp_id == COMPONENT_Cb) { count_chroma_sample_value_near_mid_point(rec_dst_yuv->u, pic_stride_c, pic_height_c, pic_width_c, max_ctu_width_log2_chrma, max_ctu_height_log2_chrma, chroma_sample_count_near_mid_point, ctus_in_width, bit_depth); } else if (comp_id == COMPONENT_Cr) { count_chroma_sample_value_near_mid_point(rec_dst_yuv->v, pic_stride_c, pic_height_c, pic_width_c, max_ctu_width_log2_chrma, max_ctu_height_log2_chrma, chroma_sample_count_near_mid_point, ctus_in_width, bit_depth); } else { assert(false); // Component ID not allowed. } }*/ for (int filter_idx = 0; filter_idx <= MAX_NUM_CC_ALF_FILTERS; filter_idx++) { if (filter_idx < MAX_NUM_CC_ALF_FILTERS) { memset(best_filter_coeff_set[filter_idx], 0, sizeof(best_filter_coeff_set[filter_idx])); best_map_filter_idx_to_filter_idc[filter_idx] = filter_idx + 1; } else { best_map_filter_idx_to_filter_idc[filter_idx] = 0; } } memset(best_filter_control, 0, sizeof(uint8_t) * num_ctus_in_pic); int cc_alf_reuse_aps_id = -1; cc_reuse_aps_id[comp_id - 1] = -1; memcpy(&ctx_start_cc_alf_filter_control_flag, cabac_estimator, sizeof(ctx_start_cc_alf_filter_control_flag)); ctx_start_cc_alf_filter_control_flag.only_count = 1; // compute cost of not filtering uint64_t unfiltered_distortion = 0; for (int ctb_idx = 0; ctb_idx < num_ctus_in_pic; ctb_idx++) { unfiltered_distortion += (uint64_t)alf_covariance_cc_alf[ctb_idx].pix_acc; } double best_unfiltered_total_cost = 1 * lambda + unfiltered_distortion; // 1 bit is for gating flag bool cc_alf_filter_idx_enabled[MAX_NUM_CC_ALF_FILTERS]; short cc_alf_filter_coeff[MAX_NUM_CC_ALF_FILTERS][MAX_NUM_CC_ALF_CHROMA_COEFF]; uint8_t cc_alf_filter_count = MAX_NUM_CC_ALF_FILTERS; double best_filtered_total_cost = MAX_DOUBLE; bool best_reuse_temporal_filter_coeff = false; int training_iter_count = 0; bool keep_training = true; bool improvement = false; double prev_total_cost = MAX_DOUBLE; const int num_coeff = CC_ALF_NUM_COEFF - 1; int log2_block_width = max_ctu_width_log2 - scale_x; int log2_block_height = max_ctu_height_log2 - scale_y; uint64_t cur_total_distortion = 0; double cur_total_rate = 0; int aps_ids_size = 0; int aps_ids[ALF_CTB_MAX_NUM_APS] = { -1, -1, -1, -1, -1, -1, -1, -1 }; get_available_cc_alf_aps_ids(state, comp_id, &aps_ids_size, aps_ids); for (int test_filter_idx = 0; test_filter_idx < (aps_ids_size + 1); test_filter_idx++) { bool referencing_existing_aps = (test_filter_idx < aps_ids_size) ? true : false; int max_number_of_filters_being_tested = MAX_NUM_CC_ALF_FILTERS - (test_filter_idx - aps_ids_size); if (max_number_of_filters_being_tested < 0) { max_number_of_filters_being_tested = 1; } { // Instead of rewriting the control buffer for every training iteration just keep a mapping from filterIdx to filterIdc uint8_t map_filter_idx_to_filter_idc[MAX_NUM_CC_ALF_FILTERS + 1]; for (int filter_idx = 0; filter_idx <= MAX_NUM_CC_ALF_FILTERS; filter_idx++) { if (filter_idx == MAX_NUM_CC_ALF_FILTERS) { map_filter_idx_to_filter_idc[filter_idx] = 0; } else { map_filter_idx_to_filter_idc[filter_idx] = filter_idx + 1; } } // initialize filters for (int filter_idx = 0; filter_idx < MAX_NUM_CC_ALF_FILTERS; filter_idx++) { cc_alf_filter_idx_enabled[filter_idx] = false; memset(cc_alf_filter_coeff[filter_idx], 0, sizeof(cc_alf_filter_coeff[filter_idx])); } if (referencing_existing_aps) { max_number_of_filters_being_tested = state->tile->frame->alf_param_set_map[aps_ids[test_filter_idx] + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set.cc_alf_aps_param.cc_alf_filter_count[comp_id - 1]; cc_alf_filter_count = max_number_of_filters_being_tested; for (int filter_idx = 0; filter_idx < max_number_of_filters_being_tested; filter_idx++) { cc_alf_filter_idx_enabled[filter_idx] = true; memcpy(cc_alf_filter_coeff[filter_idx], cc_filter_param->cc_alf_coeff[comp_id - 1][filter_idx], sizeof(cc_alf_filter_coeff[filter_idx])); } memcpy(cc_alf_filter_coeff, state->tile->frame->alf_param_set_map[aps_ids[test_filter_idx] + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set.cc_alf_aps_param.cc_alf_coeff[comp_id - 1], sizeof(cc_alf_filter_coeff)); } else { for (int i = 0; i < max_number_of_filters_being_tested; i++) { cc_alf_filter_idx_enabled[i] = true; } cc_alf_filter_count = max_number_of_filters_being_tested; } // initialize int control_idx = 0; assert(max_number_of_filters_being_tested != 0); //max_number_of_filters_being_tested should not be 0. const int column_size = (pic_width_c / max_number_of_filters_being_tested); for (int y = 0; y < pic_height_c; y += ctu_height_c) { for (int x = 0; x < pic_width_c; x += ctu_width_c) { training_cov_control[control_idx] = (x / column_size) + 1; control_idx++; } } // compute cost of filtering training_iter_count = 0; keep_training = true; improvement = false; prev_total_cost = MAX_DOUBLE; while (keep_training) { improvement = false; for (int filter_idx = 0; filter_idx < max_number_of_filters_being_tested; filter_idx++) { if (cc_alf_filter_idx_enabled[filter_idx]) { if (!referencing_existing_aps) { get_frame_stats_cc_alf(alf_covariance_cc_alf, alf_covariance_frame_cc_alf, (filter_idx + 1), num_ctus_in_pic, training_cov_control); derive_cc_alf_filter_coeff(alf_covariance_frame_cc_alf, cc_alf_filter_coeff, filter_idx); } for (int y = 0; y < pic_height_c; y += (1 << log2_block_height)) { for (int x = 0; x < pic_width_c; x += (1 << log2_block_width)) { int ctu_idx = (y >> log2_block_height) * ctus_in_width + (x >> log2_block_width); alf_info->training_distortion[filter_idx][ctu_idx] = (int)(alf_info->ctb_distortion_unfilter[comp_id][ctu_idx] + calc_error_for_cc_alf_coeffs(&alf_covariance_cc_alf[ctu_idx], cc_alf_filter_coeff[filter_idx], num_coeff, 7 + 1)); } } } } memcpy(cabac_estimator, &ctx_start_cc_alf_filter_control_flag, sizeof(*cabac_estimator)); cur_total_distortion = 0; cur_total_rate = 0; determine_control_idc_values(state, comp_id, ctu_width_c, ctu_height_c, pic_width_c, pic_height_c, alf_info->ctb_distortion_unfilter, alf_info->training_distortion, referencing_existing_aps, training_cov_control, filter_control, &cur_total_distortion, &cur_total_rate, cc_alf_filter_idx_enabled, map_filter_idx_to_filter_idc, &cc_alf_filter_count); // compute coefficient coding bit cost if (cc_alf_filter_count > 0) { if (referencing_existing_aps) { cur_total_rate += 1 + 3; // +1 for enable flag, +3 APS ID in slice header } else { cur_total_rate += get_coeff_rate_cc_alf(cc_alf_filter_coeff, cc_alf_filter_idx_enabled, cc_alf_filter_count, comp_id) + 1 + 9; // +1 for the enable flag, +9 3-bit for APS ID in slice header, 5-bit for APS ID in APS, a 1-bit // new filter flags (ignore shared cost such as other new-filter flags/NALU header/RBSP // terminating bit/byte alignment bits)*/ } double cur_total_cost = cur_total_rate * lambda + cur_total_distortion; if (cur_total_cost < prev_total_cost) { prev_total_cost = cur_total_cost; improvement = true; } if (cur_total_cost < best_filtered_total_cost) { best_filtered_total_cost = cur_total_cost; memcpy(best_filter_idx_enabled, cc_alf_filter_idx_enabled, sizeof(cc_alf_filter_idx_enabled)); memcpy(best_filter_coeff_set, cc_alf_filter_coeff, sizeof(cc_alf_filter_coeff)); memcpy(best_filter_control, filter_control, sizeof(uint8_t) * num_ctus_in_pic); best_filter_count = cc_alf_filter_count; cc_alf_reuse_aps_id = referencing_existing_aps ? aps_ids[test_filter_idx] : -1; memcpy(best_map_filter_idx_to_filter_idc, map_filter_idx_to_filter_idc, sizeof(map_filter_idx_to_filter_idc)); } } training_iter_count++; if (!improvement || training_iter_count > max_training_iter_count || referencing_existing_aps) { keep_training = false; } } } } if (best_unfiltered_total_cost < best_filtered_total_cost) { memset(best_filter_control, 0, sizeof(uint8_t) * num_ctus_in_pic); } // save best coeff and control bool atleast_one_block_undergoes_fitlering = false; for (int controlIdx = 0; best_filter_count > 0 && controlIdx < num_ctus_in_pic; controlIdx++) { if (best_filter_control[controlIdx]) { atleast_one_block_undergoes_fitlering = true; break; } } cc_filter_param->number_valid_components = (chroma_fmt == KVZ_CSP_400) ? 1 : MAX_NUM_COMPONENT; cc_filter_param->cc_alf_filter_enabled[comp_id - 1] = atleast_one_block_undergoes_fitlering; if (atleast_one_block_undergoes_fitlering) { // update the filter control indicators if (best_reuse_temporal_filter_coeff != 1) { short stored_best_filter_coeff_set[MAX_NUM_CC_ALF_FILTERS][MAX_NUM_CC_ALF_CHROMA_COEFF]; for (int filter_idx = 0; filter_idx < MAX_NUM_CC_ALF_FILTERS; filter_idx++) { memcpy(stored_best_filter_coeff_set[filter_idx], best_filter_coeff_set[filter_idx], sizeof(best_filter_coeff_set[filter_idx])); } memcpy(filter_control, best_filter_control, sizeof(uint8_t) * num_ctus_in_pic); int filter_count = 0; for (int filter_idx = 0; filter_idx < MAX_NUM_CC_ALF_FILTERS; filter_idx++) { uint8_t cur_filter_idc = best_map_filter_idx_to_filter_idc[filter_idx]; if (best_filter_idx_enabled[filter_idx]) { for (int control_idx = 0; control_idx < num_ctus_in_pic; control_idx++) { if (filter_control[control_idx] == (filter_idx + 1)) { best_filter_control[control_idx] = cur_filter_idc; } } memcpy(best_filter_coeff_set[cur_filter_idc - 1], stored_best_filter_coeff_set[filter_idx], sizeof(stored_best_filter_coeff_set[filter_idx])); filter_count++; } best_filter_idx_enabled[filter_idx] = (filter_idx < best_filter_count) ? true : false; } assert(filter_count == best_filter_count); //Number of filters enabled did not match the filter count } cc_filter_param->cc_alf_filter_count[comp_id - 1] = best_filter_count; // cleanup before copying memset(alf_info->cc_alf_filter_control[comp_id - 1], 0, sizeof(uint8_t) * num_ctus_in_pic); for (int filter_idx = 0; filter_idx < MAX_NUM_CC_ALF_FILTERS; filter_idx++) { memset(cc_filter_param->cc_alf_coeff[comp_id - 1][filter_idx], 0, sizeof(cc_filter_param->cc_alf_coeff[comp_id - 1][filter_idx])); } memset(cc_filter_param->cc_alf_filter_idx_enabled[comp_id - 1], false, sizeof(cc_filter_param->cc_alf_filter_idx_enabled[comp_id - 1])); for (int filter_idx = 0; filter_idx < best_filter_count; filter_idx++) { cc_filter_param->cc_alf_filter_idx_enabled[comp_id - 1][filter_idx] = best_filter_idx_enabled[filter_idx]; memcpy(cc_filter_param->cc_alf_coeff[comp_id - 1][filter_idx], best_filter_coeff_set[filter_idx], sizeof(best_filter_coeff_set[filter_idx])); } memcpy(alf_info->cc_alf_filter_control[comp_id - 1], best_filter_control, sizeof(uint8_t) * num_ctus_in_pic); if (cc_alf_reuse_aps_id >= 0) { cc_reuse_aps_id[comp_id - 1] = cc_alf_reuse_aps_id; if (comp_id == COMPONENT_Cb) { state->slice->alf->tile_group_cc_alf_cb_aps_id = cc_alf_reuse_aps_id; } else { state->slice->alf->tile_group_cc_alf_cr_aps_id = cc_alf_reuse_aps_id; } } } /*if (luma_swing_greater_than_threshold_count) { FREE_POINTER(luma_swing_greater_than_threshold_count); } if (chroma_sample_count_near_mid_point) { FREE_POINTER(chroma_sample_count_near_mid_point); }*/ } static void calc_covariance_cc_alf(int32_t e_local[MAX_NUM_CC_ALF_CHROMA_COEFF][1], const kvz_pixel *rec, const int stride, int vb_distance) { const kvz_pixel *rec_y_m1 = rec - 1 * stride; const kvz_pixel *rec_y_0 = rec; const kvz_pixel *rec_y_p1 = rec + 1 * stride; const kvz_pixel *rec_y_p2 = rec + 2 * stride; if (vb_distance == -2 || vb_distance == +1) { rec_y_p2 = rec_y_p1; } else if (vb_distance == -1 || vb_distance == 0) { rec_y_m1 = rec_y_0; rec_y_p2 = rec_y_p1 = rec_y_0; } const kvz_pixel center_value = rec_y_0[+0]; for (int b = 0; b < 1; b++) { e_local[0][b] += rec_y_m1[+0] - center_value; e_local[1][b] += rec_y_0[-1] - center_value; e_local[2][b] += rec_y_0[+1] - center_value; e_local[3][b] += rec_y_p1[-1] - center_value; e_local[4][b] += rec_y_p1[+0] - center_value; e_local[5][b] += rec_y_p1[+1] - center_value; e_local[6][b] += rec_y_p2[+0] - center_value; } } static void get_blk_stats_cc_alf(encoder_state_t * const state, alf_covariance *alf_covariance, const kvz_picture *org_yuv, const alf_component_id comp_id, const int x_pos, const int y_pos, const int width, const int height) { alf_info_t *alf_info = state->tile->frame->alf_info; enum kvz_chroma_format chroma_fmt = state->encoder_control->chroma_format; bool chroma_scale_x = (chroma_fmt == KVZ_CSP_444) ? 0 : 1; bool chroma_scale_y = (chroma_fmt != KVZ_CSP_420) ? 0 : 1; const int frame_height = state->tile->frame->height; const int alf_vb_luma_pos = LCU_WIDTH - ALF_VB_POS_ABOVE_CTUROW_LUMA; const int alf_vb_luma_ctu_height = LCU_WIDTH; const int max_cu_height = LCU_WIDTH; const int x_pos_c = x_pos >> chroma_scale_x; const int y_pos_c = y_pos >> chroma_scale_y; const int c_width = width >> chroma_scale_x; const int c_height = height >> chroma_scale_y; const int num_coeff = 8; const channel_type channel = (comp_id == COMPONENT_Y) ? CHANNEL_TYPE_LUMA : CHANNEL_TYPE_CHROMA; enum kvz_chroma_format chroma_format = state->encoder_control->chroma_format; const int number_of_components = (chroma_format == KVZ_CSP_400) ? 1 : MAX_NUM_COMPONENT;; int rec_stride[MAX_NUM_COMPONENT]; int rec_pixel_idx[MAX_NUM_COMPONENT]; const int luma_rec_pos = y_pos * state->tile->frame->rec->stride + x_pos; const int chroma_rec_pos = y_pos_c * (state->tile->frame->rec->stride >> chroma_scale_x) + x_pos_c; kvz_pixel *rec_y = &alf_info->alf_tmp_y[luma_rec_pos]; kvz_pixel *rec_u = &alf_info->alf_tmp_u[chroma_rec_pos]; kvz_pixel *rec_v = &alf_info->alf_tmp_v[chroma_rec_pos]; for (int c_idx = 0; c_idx < number_of_components; c_idx++) { bool is_luma = c_idx == COMPONENT_Y; rec_stride[c_idx] = state->tile->frame->rec->stride >> (is_luma ? 0 : chroma_scale_x); rec_pixel_idx[c_idx] = 0; } int org_stride = 0; const kvz_pixel *org = 0; if (comp_id == COMPONENT_Y) { org_stride = org_yuv->stride; org = &org_yuv->y[y_pos*org_stride + x_pos]; } else if (comp_id == COMPONENT_Cb) { org_stride = org_yuv->stride >> chroma_scale_x; org = &org_yuv->u[y_pos_c*org_stride + x_pos_c]; } else if (comp_id == COMPONENT_Cr) { org_stride = org_yuv->stride >> chroma_scale_x; org = &org_yuv->v[y_pos_c*org_stride + x_pos_c]; } const int num_bins = 1; int vb_ctu_height = alf_vb_luma_ctu_height; int vb_pos = alf_vb_luma_pos; if ((y_pos + max_cu_height) >= frame_height) { vb_pos = frame_height; } int32_t e_local[MAX_NUM_CC_ALF_CHROMA_COEFF][1]; kvz_pixel *rec_pixels = (comp_id == COMPONENT_Y ? rec_y : (comp_id == COMPONENT_Cb ? rec_u : rec_v)); uint8_t component_scale_y = (comp_id == COMPONENT_Y || chroma_format != KVZ_CSP_420) ? 0 : 1; uint8_t component_scale_x = (comp_id == COMPONENT_Y || chroma_format == KVZ_CSP_444) ? 0 : 1; int16_t y_local = 0; for (int i = 0; i < (comp_id == COMPONENT_Y ? height : c_height); i++) { int vb_distance = ((i << component_scale_y) % vb_ctu_height) - vb_pos; const bool skip_this_row = (component_scale_y == 0 && (vb_distance == 0 || vb_distance == 1)); for (int j = 0; j < (comp_id == COMPONENT_Y ? width : c_width) && (!skip_this_row); j++) { memset(e_local, 0, sizeof(e_local)); double weight = 1.0; if (0 /*g_alf_wssd*/) { //weight = m_lumaLevelToWeightPLUT[org[j]]; } y_local = org[j] - rec_pixels[j + rec_pixel_idx[comp_id]]; calc_covariance_cc_alf(e_local, rec_y + rec_pixel_idx[COMPONENT_Y] + (j << component_scale_x), rec_stride[COMPONENT_Y], vb_distance); for (int k = 0; k < (num_coeff - 1); k++) { for (int l = k; l < (num_coeff - 1); l++) { for (int b0 = 0; b0 < num_bins; b0++) { for (int b1 = 0; b1 < num_bins; b1++) { if (0 /*g_alf_wssd*/) { alf_covariance->ee[b0][b1][k][l] += weight * (e_local[k][b0] * (double)e_local[l][b1]); } else { alf_covariance->ee[b0][b1][k][l] += e_local[k][b0] * (double)e_local[l][b1]; } } } } for (int b = 0; b < num_bins; b++) { if (0 /*g_alf_wssd*/) { alf_covariance->y[b][k] += weight * (e_local[k][b] * (double)y_local); } else { alf_covariance->y[b][k] += e_local[k][b] * (double)y_local; } } } if (0 /*g_alf_wssd*/) { alf_covariance->pix_acc += weight * (y_local * (double)y_local); } else { alf_covariance->pix_acc += y_local * (double)y_local; } } org += org_stride; for (int src_c_idx = 0; src_c_idx < number_of_components; src_c_idx++) { const channel_type c_channel = (src_c_idx == COMPONENT_Y) ? CHANNEL_TYPE_LUMA : CHANNEL_TYPE_CHROMA; if (c_channel == channel) { rec_pixel_idx[src_c_idx] += rec_stride[src_c_idx]; } else { if (comp_id == COMPONENT_Y) { rec_pixel_idx[src_c_idx] += rec_stride[src_c_idx] >> ((src_c_idx == COMPONENT_Y || chroma_format != KVZ_CSP_420) ? 0 : 1); } else { rec_pixel_idx[src_c_idx] += rec_stride[src_c_idx] << ((comp_id == COMPONENT_Y || chroma_format != KVZ_CSP_420) ? 0 : 1); } } } } for (int k = 1; k < (MAX_NUM_CC_ALF_CHROMA_COEFF - 1); k++) { for (int l = 0; l < k; l++) { for (int b0 = 0; b0 < num_bins; b0++) { for (int b1 = 0; b1 < num_bins; b1++) { alf_covariance->ee[b0][b1][k][l] = alf_covariance->ee[b1][b0][l][k]; } } } } } static void derive_stats_for_cc_alf_filtering(encoder_state_t * const state, const kvz_picture *org_yuv, const int comp_idx, const int mask_stride, const uint8_t filter_idc) { alf_covariance **alf_covariance_cc_alf = state->tile->frame->alf_info->alf_covariance_cc_alf; alf_covariance *alf_covariance_frame_cc_alf = state->tile->frame->alf_info->alf_covariance_frame_cc_alf[comp_idx - 1]; const int32_t num_ctus_in_pic = state->tile->frame->width_in_lcu * state->tile->frame->height_in_lcu; const int filter_idx = filter_idc - 1; // init CTU stats buffers for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { reset_alf_covariance(&alf_covariance_cc_alf[comp_idx - 1][(filter_idx * num_ctus_in_pic) + ctu_idx], -1); } // init Frame stats buffers reset_alf_covariance(&alf_covariance_frame_cc_alf[filter_idx], -1); int ctu_rs_addr = 0; const int frame_height = state->tile->frame->height; const int frame_width = state->tile->frame->width; const int max_cu_width = LCU_WIDTH; const int max_cu_height = LCU_WIDTH; for (int y_pos = 0; y_pos < state->tile->frame->height; y_pos += max_cu_height) { for (int x_pos = 0; x_pos < state->tile->frame->width; x_pos += max_cu_width) { const int width = (x_pos + max_cu_width > frame_width) ? (frame_width - x_pos) : max_cu_width; const int height = (y_pos + max_cu_height > frame_height) ? (frame_height - y_pos) : max_cu_height; get_blk_stats_cc_alf(state, &alf_covariance_cc_alf[comp_idx - 1][(filter_idx * num_ctus_in_pic) + ctu_rs_addr], org_yuv, comp_idx, x_pos, y_pos, width, height); add_alf_cov(&alf_covariance_frame_cc_alf[filter_idx], &alf_covariance_cc_alf[comp_idx - 1][(filter_idx * num_ctus_in_pic) + ctu_rs_addr]); ctu_rs_addr++; } } } static void count_luma_swing_greater_than_threshold(const kvz_pixel* luma, int luma_stride, int height, int width, int log2_block_width, int log2_block_height, uint64_t* luma_swing_greater_than_threshold_count, int luma_count_stride, int8_t input_bit_depth) { const int threshold = (1 << (input_bit_depth - 2)) - 1; // 3x4 Diamond int x_support[] = { 0, -1, 0, 1, -1, 0, 1, 0 }; int y_support[] = { -1, 0, 0, 0, 1, 1, 1, 2 }; for (int y = 0; y < height; y += (1 << log2_block_height)) { for (int x = 0; x < width; x += (1 << log2_block_width)) { luma_swing_greater_than_threshold_count[(y >> log2_block_height) * luma_count_stride + (x >> log2_block_width)] = 0; for (int y_off = 0; y_off < (1 << log2_block_height); y_off++) { for (int x_off = 0; x_off < (1 << log2_block_width); x_off++) { if ((y + y_off) >= (height - 2) || (x + x_off) >= (width - 1) || (y + y_off) < 1 || (x + x_off) < 1) // only consider samples that are fully supported by picture { continue; } int min_val = ((1 << input_bit_depth) - 1); int max_val = 0; for (int i = 0; i < 8; i++) { kvz_pixel p = luma[(y_off + y_support[i]) * luma_stride + x + x_off + x_support[i]]; if (p < min_val) { min_val = p; } if (p > max_val) { max_val = p; } } if ((max_val - min_val) > threshold) { luma_swing_greater_than_threshold_count[(y >> log2_block_height) * luma_count_stride + (x >> log2_block_width)]++; } } } } luma += (luma_stride << log2_block_height); } } static void count_chroma_sample_value_near_mid_point(const kvz_pixel* chroma, int chroma_stride, int height, int width, int log2_block_width, int log2_block_height, uint64_t* chroma_sample_count_near_mid_point, int chroma_sample_count_near_mid_point_stride, int8_t input_bit_depth) { const int mid_point = (1 << input_bit_depth) >> 1; const int threshold = 16; for (int y = 0; y < height; y += (1 << log2_block_height)) { for (int x = 0; x < width; x += (1 << log2_block_width)) { chroma_sample_count_near_mid_point[(y >> log2_block_height) * chroma_sample_count_near_mid_point_stride + (x >> log2_block_width)] = 0; for (int y_off = 0; y_off < (1 << log2_block_height); y_off++) { for (int x_off = 0; x_off < (1 << log2_block_width); x_off++) { if ((y + y_off) >= height || (x + x_off) >= width) { continue; } int distance_to_mid_point = abs(chroma[y_off * chroma_stride + x + x_off] - mid_point); if (distance_to_mid_point < threshold) { chroma_sample_count_near_mid_point[(y >> log2_block_height) * chroma_sample_count_near_mid_point_stride + (x >> log2_block_width)]++; } } } } chroma += (chroma_stride << log2_block_height); } } static void init_distortion_cc_alf(alf_covariance* alf_covariance_cc_alf[MAX_NUM_COMPONENT], double **ctb_distortion_unfilter, const int num_ctus) { for (int comp = 1; comp < MAX_NUM_COMPONENT; comp++) { for (int ctb_idx = 0; ctb_idx < num_ctus; ctb_idx++) { ctb_distortion_unfilter[comp][ctb_idx] = alf_covariance_cc_alf[comp - 1][ctb_idx].pix_acc; } } } static void alf_reconstruct_coeff(encoder_state_t * const state, alf_aps *aps, channel_type channel, const bool is_rdo, const bool is_redo, array_variables *arr_vars) { const int8_t bit_depth = state->encoder_control->bitdepth; int factor = is_rdo ? 0 : (1 << (bit_depth - 1)); bool is_luma = channel == CHANNEL_TYPE_LUMA ? 1 : 0; alf_filter_type filter_type = is_luma ? ALF_FILTER_7X7 : ALF_FILTER_5X5; int num_classes = is_luma ? MAX_NUM_ALF_CLASSES : 1; int num_coeff = filter_type == ALF_FILTER_5X5 ? 7 : 13; int num_coeff_minus1 = num_coeff - 1; const int num_alts = is_luma ? 1 : aps->num_alternatives_chroma; for (int alt_idx = 0; alt_idx < num_alts; ++alt_idx) { int num_filters = is_luma ? aps->num_luma_filters : 1; short* coeff = is_luma ? aps->luma_coeff : aps->chroma_coeff[alt_idx]; int16_t* clipp = is_luma ? aps->luma_clipp : aps->chroma_clipp[alt_idx]; for (int filter_idx = 0; filter_idx < num_filters; filter_idx++) { coeff[filter_idx * MAX_NUM_ALF_LUMA_COEFF + num_coeff_minus1] = factor; } if (!is_luma) { for (int coeff_idx = 0; coeff_idx < num_coeff_minus1; ++coeff_idx) { arr_vars->chroma_coeff_final[alt_idx][coeff_idx] = coeff[coeff_idx]; int clip_idx = aps->non_linear_flag[channel] ? clipp[coeff_idx] : 0; arr_vars->chroma_clipp_final[alt_idx][coeff_idx] = is_rdo ? clip_idx : arr_vars->alf_clipping_values[channel][clip_idx]; } arr_vars->chroma_coeff_final[alt_idx][num_coeff_minus1] = factor; arr_vars->chroma_clipp_final[alt_idx][num_coeff_minus1] = is_rdo ? 0 : arr_vars->alf_clipping_values[channel][0]; continue; } for (int class_idx = 0; class_idx < num_classes; class_idx++) { int filter_idx = aps->filter_coeff_delta_idx[class_idx]; assert((filter_idx >= 0 && filter_idx <= aps->num_luma_filters)); // "Bad coeff delta idx in ALF" for (int coeff_idx = 0; coeff_idx < num_coeff_minus1; ++coeff_idx) { arr_vars->coeff_final[class_idx * MAX_NUM_ALF_LUMA_COEFF + coeff_idx] = coeff[filter_idx * MAX_NUM_ALF_LUMA_COEFF + coeff_idx]; } arr_vars->coeff_final[class_idx* MAX_NUM_ALF_LUMA_COEFF + num_coeff_minus1] = factor; arr_vars->clipp_final[class_idx* MAX_NUM_ALF_LUMA_COEFF + num_coeff_minus1] = is_rdo ? 0 : arr_vars->alf_clipping_values[channel][0]; for (int coeff_idx = 0; coeff_idx < num_coeff_minus1; ++coeff_idx) { int clip_idx = aps->non_linear_flag[channel] ? clipp[filter_idx * MAX_NUM_ALF_LUMA_COEFF + coeff_idx] : 0; assert((clip_idx >= 0 && clip_idx < MAX_ALF_NUM_CLIPPING_VALUES)); // "Bad clip idx in ALF" arr_vars->clipp_final[class_idx * MAX_NUM_ALF_LUMA_COEFF + coeff_idx] = is_rdo ? clip_idx : arr_vars->alf_clipping_values[channel][clip_idx]; } arr_vars->clipp_final[class_idx* MAX_NUM_ALF_LUMA_COEFF + num_coeff_minus1] = is_rdo ? 0 : arr_vars->alf_clipping_values[channel][0]; } } } static double alf_derive_ctb_alf_enable_flags(encoder_state_t * const state, channel_type channel, double *dist_unfilter, const int num_classes, const double chroma_weight, array_variables *arr_vars) { alf_info_t *alf_info = state->tile->frame->alf_info; short* alf_ctb_filter_index = alf_info->alf_ctb_filter_index; cabac_data_t *cabac_estimator = &alf_info->cabac_estimator; cabac_data_t ctx_temp_start; cabac_data_t ctx_temp_best; cabac_data_t ctx_temp_alt_start; //cabac_data_t ctx_temp_alt_best; bool is_luma = channel == CHANNEL_TYPE_LUMA ? 1 : 0; alf_aps *alf_param_temp = &alf_info->alf_param_temp; const kvz_pixel comp_id_first = is_luma ? COMPONENT_Y : COMPONENT_Cb; const kvz_pixel comp_id_last = is_luma ? COMPONENT_Y : COMPONENT_Cr; const int num_alts = is_luma ? 1 : alf_param_temp->num_alternatives_chroma; const int8_t bit_depth = state->encoder_control->bitdepth; const int32_t num_ctus_in_pic = state->tile->frame->width_in_lcu * state->tile->frame->height_in_lcu; int num_coeff = is_luma ? 13 : 7; double cost = 0; double lambda = state->frame->lambda; *dist_unfilter = 0; if (is_luma) { alf_param_temp->enabled_flag[COMPONENT_Y] = 1; } else { alf_param_temp->enabled_flag[COMPONENT_Cb] = 1; alf_param_temp->enabled_flag[COMPONENT_Cr] = 1; } assert((chroma_weight <= 0.0) && (state->slice->start_in_rs == 0)); //"incompatible start CTU address, must be 0" alf_reconstruct_coeff(state, alf_param_temp, channel, true, is_luma, arr_vars); for (int alt_idx = 0; alt_idx < (is_luma ? 1 : MAX_NUM_ALF_ALTERNATIVES_CHROMA); alt_idx++) { for (int class_idx = 0; class_idx < (is_luma ? MAX_NUM_ALF_CLASSES : 1); class_idx++) { for (int i = 0; i < (is_luma ? MAX_NUM_ALF_LUMA_COEFF : MAX_NUM_ALF_CHROMA_COEFF); i++) { arr_vars->filter_coeff_set[is_luma ? class_idx : alt_idx][i] = is_luma ? arr_vars->coeff_final[class_idx * MAX_NUM_ALF_LUMA_COEFF + i] : arr_vars->chroma_coeff_final[alt_idx][i]; arr_vars->filter_clipp_set[is_luma ? class_idx : alt_idx][i] = is_luma ? arr_vars->clipp_final[class_idx * MAX_NUM_ALF_LUMA_COEFF + i] : arr_vars->chroma_clipp_final[alt_idx][i]; } } } alf_covariance* alf_cov; for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { for (int comp_id = comp_id_first; comp_id <= comp_id_last; comp_id++) { alf_cov = (is_luma ? alf_info->alf_covariance_y : (comp_id == COMPONENT_Cb ? alf_info->alf_covariance_u : alf_info->alf_covariance_v)); const double ctu_lambda = chroma_weight > 0.0 ? (is_luma ? 0/*cs.picture->m_uEnerHpCtu[ctuIdx]*/ : 0/*cs.picture->m_uEnerHpCtu[ctuIdx]*/ / chroma_weight) : lambda; double dist_unfilter_ctu = get_unfiltered_distortion_cov_classes(&alf_cov[ctu_idx * num_classes], num_classes); //ctxTempStart = AlfCtx(m_CABACEstimator->getCtx()); memcpy(&ctx_temp_start, cabac_estimator, sizeof(ctx_temp_start)); //m_CABACEstimator->resetBits(); alf_cabac_reset_bits(cabac_estimator); cabac_estimator->only_count = 1; alf_info->ctu_enable_flag[comp_id][ctu_idx] = 1; code_alf_ctu_enable_flag(state, cabac_estimator, ctu_idx, comp_id, alf_param_temp); if (is_luma) { // Evaluate cost of signaling filter set index for convergence of filters enabled flag / filter derivation assert(alf_ctb_filter_index[ctu_idx] == ALF_NUM_FIXED_FILTER_SETS); assert(state->slice->alf->tile_group_num_aps == 1); code_alf_ctu_filter_index(state, cabac_estimator, ctu_idx, alf_param_temp->enabled_flag[COMPONENT_Y]); } double cost_on = dist_unfilter_ctu + ctu_lambda * (23 - cabac_estimator->bits_left) + (cabac_estimator->num_buffered_bytes << 3); //ctxTempBest = AlfCtx(m_CABACEstimator->getCtx()); memcpy(&ctx_temp_best, cabac_estimator, sizeof(ctx_temp_best)); if (is_luma) { cost_on += get_filtered_distortion(&alf_cov[ctu_idx * num_classes], arr_vars, num_classes, alf_param_temp->num_luma_filters - 1, num_coeff, bit_depth); } else { double best_alt_cost = MAX_DOUBLE; int best_alt_idx = -1; //ctxTempAltStart = AlfCtx(ctxTempBest); memcpy(&ctx_temp_alt_start, &ctx_temp_best, sizeof(ctx_temp_alt_start)); for (int alt_idx = 0; alt_idx < num_alts; ++alt_idx) { if (alt_idx) { //m_CABACEstimator->getCtx() = AlfCtx(ctxTempAltStart); memcpy(cabac_estimator, &ctx_temp_alt_start, sizeof(*cabac_estimator)); } //m_CABACEstimator->resetBits(); alf_cabac_reset_bits(cabac_estimator); cabac_estimator->only_count = 1; alf_info->ctu_alternative[comp_id][ctu_idx] = alt_idx; code_alf_ctu_alternative_ctu(state, cabac_estimator, ctu_idx, comp_id, alf_param_temp); double r_altCost = ctu_lambda * (23 - cabac_estimator->bits_left) + (cabac_estimator->num_buffered_bytes << 3); //frac_bits_scale * 0/*m_CABACEstimator->getEstFracBits()*/; double alt_dist = 0.; alt_dist += calc_error_for_coeffs(&alf_cov[ctu_idx * num_classes], arr_vars->filter_clipp_set[alt_idx], arr_vars->filter_coeff_set[alt_idx], num_coeff, bit_depth); double alt_cost = alt_dist + r_altCost; if (alt_cost < best_alt_cost) { best_alt_cost = alt_cost; best_alt_idx = alt_idx; //ctxTempBest = AlfCtx(m_CABACEstimator->getCtx()); memcpy(&ctx_temp_best, cabac_estimator, sizeof(ctx_temp_best)); } } alf_info->ctu_alternative[comp_id][ctu_idx] = best_alt_idx; cost_on += best_alt_cost; } //m_CABACEstimator->getCtx() = AlfCtx(ctxTempStart); memcpy(cabac_estimator, &ctx_temp_start, sizeof(*cabac_estimator)); //m_CABACEstimator->resetBits(); alf_cabac_reset_bits(cabac_estimator); cabac_estimator->only_count = 1; alf_info->ctu_enable_flag[comp_id][ctu_idx] = 0; code_alf_ctu_enable_flag(state, cabac_estimator, ctu_idx, comp_id, alf_param_temp); double cost_off = dist_unfilter_ctu + ctu_lambda * (23 - cabac_estimator->bits_left) + (cabac_estimator->num_buffered_bytes << 3); //frac_bits_scale * 0;// m_CABACEstimator->getEstFracBits(); if (cost_on < cost_off) { cost += cost_on; //m_CABACEstimator->getCtx() = AlfCtx(ctxTempBest); memcpy(cabac_estimator, &ctx_temp_best, sizeof(*cabac_estimator)); alf_info->ctu_enable_flag[comp_id][ctu_idx] = 1; } else { cost += cost_off; alf_info->ctu_enable_flag[comp_id][ctu_idx] = 0; *dist_unfilter += dist_unfilter_ctu; } } } if (!is_luma) { const alf_component_id compIDFirst = COMPONENT_Cb; const alf_component_id compIDLast = COMPONENT_Cr; for (int compId = compIDFirst; compId <= compIDLast; compId++) { alf_param_temp->enabled_flag[compId] = false; for (int i = 0; i < num_ctus_in_pic; i++) { if (alf_info->ctu_enable_flag[compId][i]) { alf_param_temp->enabled_flag[compId] = true; break; } } } } return cost; } static void alf_create_frame_buffer(encoder_state_t * const state, alf_info_t *alf_info) { if (!alf_info->alf_fulldata_buf) { enum kvz_chroma_format chroma_format = state->encoder_control->chroma_format; const size_t simd_padding_width = 64; int width = state->tile->frame->width; int height = state->tile->frame->height; int stride = state->tile->frame->source->stride; unsigned int luma_size = (width + 8) * (height + 8); unsigned chroma_sizes[] = { 0, luma_size / 4, luma_size / 2, luma_size }; unsigned chroma_size = chroma_sizes[chroma_format]; alf_info->alf_fulldata_buf = MALLOC_SIMD_PADDED(kvz_pixel, (luma_size + 2 * chroma_size), simd_padding_width * 2); alf_info->alf_fulldata = &alf_info->alf_fulldata_buf[4 * (width + 8) + 4] + simd_padding_width / sizeof(kvz_pixel); alf_info->alf_tmp_y = &alf_info->alf_fulldata[0]; if (chroma_format == KVZ_CSP_400) { alf_info->alf_tmp_u = NULL; alf_info->alf_tmp_v = NULL; } else { alf_info->alf_tmp_u = &alf_info->alf_fulldata[luma_size - (4 * (width + 8) + 4) + (2 * (stride / 2) + 2)]; alf_info->alf_tmp_v = &alf_info->alf_fulldata[luma_size - (4 * (width + 8) + 4) + chroma_size + (2 * (stride / 2) + 2)]; } } } static void alf_init_covariance(videoframe_t* frame, enum kvz_chroma_format chroma_format) { const int num_ctus_in_pic = frame->width_in_lcu * frame->height_in_lcu; const int pic_width = frame->width; const int pic_height = frame->height; const int luma_coeffs = 13; const int chroma_coeffs = 7; const int cc_alf_coeff = 8; int num_classes = 0; if (chroma_format != KVZ_CSP_400) { for (int comp_idx = 0; comp_idx < MAX_NUM_COMPONENT; comp_idx++) { num_classes += comp_idx ? 1 : MAX_NUM_ALF_CLASSES; } } else { num_classes = MAX_NUM_ALF_CLASSES; } alf_info_t* alf_info = frame->alf_info; const int num_covs = num_ctus_in_pic * num_classes; const int num_luma_covs = num_ctus_in_pic * MAX_NUM_ALF_CLASSES; alf_info->alf_covariance = malloc(num_covs * sizeof(alf_covariance)); alf_info->alf_covariance_y = &alf_info->alf_covariance[0]; for (int indx = 0; indx < num_luma_covs; indx++) { init_alf_covariance(&frame->alf_info->alf_covariance_y[indx], luma_coeffs); } for (int k = 0; k < MAX_NUM_ALF_CLASSES; k++) { init_alf_covariance(&alf_info->alf_covariance_frame_luma[k], luma_coeffs); } if (chroma_format != KVZ_CSP_400) { const int num_chroma_covs = num_ctus_in_pic; alf_info->alf_covariance_u = &alf_info->alf_covariance[num_luma_covs]; alf_info->alf_covariance_v = &alf_info->alf_covariance[num_luma_covs + num_chroma_covs]; for (int k = 0; k < num_chroma_covs; k++) { init_alf_covariance(&alf_info->alf_covariance_u[k], chroma_coeffs); init_alf_covariance(&alf_info->alf_covariance_v[k], chroma_coeffs); } for (int k = 0; k < MAX_NUM_ALF_ALTERNATIVES_CHROMA; k++) { init_alf_covariance(&alf_info->alf_covariance_frame_chroma[k], chroma_coeffs); } alf_info->alf_covariance_cc_alf[MAX_NUM_COMPONENT - 1] = malloc(num_ctus_in_pic * MAX_NUM_CC_ALF_FILTERS * (MAX_NUM_COMPONENT - 1) * sizeof(alf_covariance)); for (int comp_idx = 0; comp_idx < (MAX_NUM_COMPONENT - 1); comp_idx++) { alf_info->alf_covariance_cc_alf[comp_idx] = &alf_info->alf_covariance_cc_alf[MAX_NUM_COMPONENT - 1][comp_idx * MAX_NUM_CC_ALF_FILTERS * num_ctus_in_pic]; } for (int k = 0; k < num_ctus_in_pic * MAX_NUM_CC_ALF_FILTERS * (MAX_NUM_COMPONENT - 1); k++) { init_alf_covariance(&alf_info->alf_covariance_cc_alf[MAX_NUM_COMPONENT - 1][k], cc_alf_coeff); } for (int comp_idx = 0; comp_idx < MAX_NUM_COMPONENT - 1; comp_idx++) { for (int k = 0; k < MAX_NUM_CC_ALF_FILTERS; k++) { init_alf_covariance(&alf_info->alf_covariance_frame_cc_alf[comp_idx][k], cc_alf_coeff); } } } for (int k = 0; k <= MAX_NUM_ALF_CLASSES + 1; k++) { init_alf_covariance(&alf_info->alf_covariance_merged[k], luma_coeffs); } alf_info->training_cov_control = malloc(num_ctus_in_pic * sizeof(*alf_info->training_cov_control)); alf_info->training_distortion[MAX_NUM_CC_ALF_FILTERS] = malloc(num_ctus_in_pic * MAX_NUM_CC_ALF_FILTERS * sizeof(*alf_info->training_distortion[MAX_NUM_CC_ALF_FILTERS])); memset(alf_info->training_distortion[MAX_NUM_CC_ALF_FILTERS], 0, num_ctus_in_pic * MAX_NUM_CC_ALF_FILTERS * sizeof(*alf_info->training_distortion[MAX_NUM_CC_ALF_FILTERS])); for (int i = 0; i < MAX_NUM_CC_ALF_FILTERS; i++) { alf_info->training_distortion[i] = &alf_info->training_distortion[MAX_NUM_CC_ALF_FILTERS][num_ctus_in_pic * i]; } alf_info->filter_control = malloc(num_ctus_in_pic * sizeof(*alf_info->filter_control)); alf_info->best_filter_control = malloc(num_ctus_in_pic * sizeof(*alf_info->best_filter_control)); // Classification alf_info->classifier = malloc(pic_height * sizeof(alf_classifier*)); alf_info->classifier[0] = malloc(pic_height * pic_width * sizeof(alf_classifier)); for (int i = 1; i < pic_height; i++) { alf_info->classifier[i] = alf_info->classifier[0] + i * pic_width; } } void kvz_alf_create(videoframe_t *frame, enum kvz_chroma_format chroma_format) { const int num_ctus_in_pic = frame->width_in_lcu * frame->height_in_lcu; const int pic_width = frame->width; const int pic_height = frame->height; const int luma_coeffs = 13; const int chroma_coeffs = 7; const int cc_alf_coeff = 8; int num_classes = 0; alf_info_t *alf_info = frame->alf_info; alf_info->aps_id_start = ALF_CTB_MAX_NUM_APS; alf_info->ctu_enable_flag[MAX_NUM_COMPONENT] = malloc(num_ctus_in_pic * MAX_NUM_COMPONENT * sizeof(*alf_info->ctu_enable_flag[MAX_NUM_COMPONENT])); memset(alf_info->ctu_enable_flag[MAX_NUM_COMPONENT], 0, num_ctus_in_pic * MAX_NUM_COMPONENT * sizeof(*alf_info->ctu_enable_flag[MAX_NUM_COMPONENT])); alf_info->ctu_enable_flag_tmp[MAX_NUM_COMPONENT] = malloc(num_ctus_in_pic * MAX_NUM_COMPONENT * sizeof(*alf_info->ctu_enable_flag_tmp[MAX_NUM_COMPONENT])); memset(alf_info->ctu_enable_flag_tmp[MAX_NUM_COMPONENT], 0, num_ctus_in_pic * MAX_NUM_COMPONENT * sizeof(*alf_info->ctu_enable_flag_tmp[MAX_NUM_COMPONENT])); alf_info->ctu_alternative[MAX_NUM_COMPONENT] = malloc(num_ctus_in_pic * (MAX_NUM_COMPONENT - 1) * sizeof(*alf_info->ctu_alternative[MAX_NUM_COMPONENT])); memset(alf_info->ctu_alternative[MAX_NUM_COMPONENT], 0, num_ctus_in_pic * (MAX_NUM_COMPONENT - 1) * sizeof(*alf_info->ctu_alternative[MAX_NUM_COMPONENT])); alf_info->ctu_alternative_tmp[MAX_NUM_COMPONENT] = malloc(num_ctus_in_pic * (MAX_NUM_COMPONENT - 1) * sizeof(*alf_info->ctu_alternative_tmp[MAX_NUM_COMPONENT])); memset(alf_info->ctu_alternative_tmp[MAX_NUM_COMPONENT], 0, num_ctus_in_pic * (MAX_NUM_COMPONENT - 1) * sizeof(*alf_info->ctu_alternative_tmp[MAX_NUM_COMPONENT])); alf_info->ctb_distortion_unfilter[MAX_NUM_COMPONENT] = malloc(num_ctus_in_pic * MAX_NUM_COMPONENT * sizeof(*alf_info->ctb_distortion_unfilter[MAX_NUM_COMPONENT])); memset(alf_info->ctb_distortion_unfilter[MAX_NUM_COMPONENT], 0, num_ctus_in_pic * MAX_NUM_COMPONENT * sizeof(*alf_info->ctb_distortion_unfilter[MAX_NUM_COMPONENT])); for (int comp_idx = 0; comp_idx < MAX_NUM_COMPONENT; comp_idx++) { alf_info->ctu_enable_flag[comp_idx] = &alf_info->ctu_enable_flag[MAX_NUM_COMPONENT][comp_idx * num_ctus_in_pic]; alf_info->ctu_enable_flag_tmp[comp_idx] = &alf_info->ctu_enable_flag_tmp[MAX_NUM_COMPONENT][comp_idx * num_ctus_in_pic]; alf_info->ctb_distortion_unfilter[comp_idx] = &alf_info->ctb_distortion_unfilter[MAX_NUM_COMPONENT][comp_idx * num_ctus_in_pic]; if (comp_idx == COMPONENT_Y) { alf_info->ctu_alternative[comp_idx] = NULL; alf_info->ctu_alternative_tmp[comp_idx] = NULL; } else { alf_info->ctu_alternative[comp_idx] = &alf_info->ctu_alternative[MAX_NUM_COMPONENT][(comp_idx - 1) * num_ctus_in_pic]; alf_info->ctu_alternative_tmp[comp_idx] = &alf_info->ctu_alternative_tmp[MAX_NUM_COMPONENT][(comp_idx - 1) * num_ctus_in_pic]; } } if (chroma_format != KVZ_CSP_400) { for (int comp_idx = 0; comp_idx < MAX_NUM_COMPONENT; comp_idx++) { num_classes += comp_idx ? 1 : MAX_NUM_ALF_CLASSES; } } else { num_classes = MAX_NUM_ALF_CLASSES; } alf_info->cc_alf_filter_control[2] = malloc(2 * num_ctus_in_pic * sizeof(*alf_info->cc_alf_filter_control[2])); memset(alf_info->cc_alf_filter_control[2], 0, 2 * num_ctus_in_pic * sizeof(*alf_info->cc_alf_filter_control[2])); alf_info->cc_alf_filter_control[0] = &alf_info->cc_alf_filter_control[2][0]; alf_info->cc_alf_filter_control[1] = &alf_info->cc_alf_filter_control[2][num_ctus_in_pic]; alf_info->alf_ctb_filter_index = malloc(num_ctus_in_pic * sizeof(*alf_info->alf_ctb_filter_index)); alf_info->alf_ctb_filter_set_index_tmp = malloc(num_ctus_in_pic * sizeof(*alf_info->alf_ctb_filter_set_index_tmp)); alf_info->alf_fulldata_buf = NULL; alf_info->alf_fulldata = NULL; alf_info->alf_tmp_y = NULL; alf_info->alf_tmp_u = NULL; alf_info->alf_tmp_v = NULL; } static void alf_covariance_destroy(videoframe_t* const frame) { alf_info_t* alf_info = frame->alf_info; if (alf_info->alf_covariance_y) { alf_info->alf_covariance_y = NULL; } if (alf_info->alf_covariance_u) { alf_info->alf_covariance_u = NULL; } if (alf_info->alf_covariance_v) { alf_info->alf_covariance_v = NULL; } if (alf_info->alf_covariance) { FREE_POINTER(alf_info->alf_covariance); } for (int comp_idx = 0; comp_idx < MAX_NUM_COMPONENT; comp_idx++) { if (comp_idx > 0) { if (alf_info->alf_covariance_cc_alf[comp_idx - 1]) { alf_info->alf_covariance_cc_alf[comp_idx - 1] = NULL; } } } if (alf_info->classifier) { FREE_POINTER(alf_info->classifier[0]); FREE_POINTER(alf_info->classifier); } if (alf_info->training_cov_control) { FREE_POINTER(alf_info->training_cov_control); } for (int i = 0; i < MAX_NUM_CC_ALF_FILTERS; i++) { if (alf_info->training_distortion[i]) { alf_info->training_distortion[i] = NULL; } } if (alf_info->training_distortion[MAX_NUM_CC_ALF_FILTERS]) { FREE_POINTER(alf_info->training_distortion[MAX_NUM_CC_ALF_FILTERS]); } if (alf_info->filter_control) { FREE_POINTER(alf_info->filter_control); } if (alf_info->best_filter_control) { FREE_POINTER(alf_info->best_filter_control); } if (alf_info->alf_covariance_cc_alf[MAX_NUM_COMPONENT - 1]) { FREE_POINTER(alf_info->alf_covariance_cc_alf[MAX_NUM_COMPONENT - 1]); } } void kvz_alf_destroy(videoframe_t * const frame) { alf_info_t *alf_info = frame->alf_info; for (int comp_idx = 0; comp_idx < MAX_NUM_COMPONENT; comp_idx++) { if (alf_info->ctu_enable_flag[comp_idx]) { alf_info->ctu_enable_flag[comp_idx] = NULL; } if (alf_info->ctu_enable_flag_tmp[comp_idx]) { alf_info->ctu_enable_flag_tmp[comp_idx] = NULL; } if (alf_info->ctu_alternative[comp_idx]) { alf_info->ctu_alternative[comp_idx] = NULL; } if (alf_info->ctu_alternative_tmp[comp_idx]) { alf_info->ctu_alternative_tmp[comp_idx] = NULL; } if (alf_info->ctb_distortion_unfilter[comp_idx]) { alf_info->ctb_distortion_unfilter[comp_idx] = NULL; } } if (alf_info->ctu_enable_flag[MAX_NUM_COMPONENT]) { FREE_POINTER(alf_info->ctu_enable_flag[MAX_NUM_COMPONENT]); } if (alf_info->ctu_enable_flag_tmp[MAX_NUM_COMPONENT]) { FREE_POINTER(alf_info->ctu_enable_flag_tmp[MAX_NUM_COMPONENT]); } if (alf_info->ctu_alternative[MAX_NUM_COMPONENT]) { FREE_POINTER(alf_info->ctu_alternative[MAX_NUM_COMPONENT]); } if (alf_info->ctu_alternative_tmp[MAX_NUM_COMPONENT]) { FREE_POINTER(alf_info->ctu_alternative_tmp[MAX_NUM_COMPONENT]); } if (alf_info->ctb_distortion_unfilter[MAX_NUM_COMPONENT]) { FREE_POINTER(alf_info->ctb_distortion_unfilter[MAX_NUM_COMPONENT]); } if (alf_info->cc_alf_filter_control[0]) { alf_info->cc_alf_filter_control[0] = NULL; } if (alf_info->cc_alf_filter_control[1]) { alf_info->cc_alf_filter_control[1] = NULL; } if (alf_info->cc_alf_filter_control[2]) { FREE_POINTER(alf_info->cc_alf_filter_control[2]); } if (alf_info->alf_ctb_filter_index) { FREE_POINTER(alf_info->alf_ctb_filter_index); } if (alf_info->alf_ctb_filter_set_index_tmp) { FREE_POINTER(alf_info->alf_ctb_filter_set_index_tmp); } if (alf_info->alf_tmp_y) { alf_info->alf_tmp_y = NULL; } if (alf_info->alf_tmp_u) { alf_info->alf_tmp_u = NULL; } if (alf_info->alf_tmp_v) { alf_info->alf_tmp_v = NULL; } if (alf_info->alf_fulldata) { alf_info->alf_fulldata = NULL; } if (alf_info->alf_fulldata_buf) { FREE_POINTER(alf_info->alf_fulldata_buf); } } static void alf_merge_classes(alf_aps *alf_aps, channel_type channel, alf_covariance* cov, alf_covariance* cov_merged, int clip_merged[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF], const int num_classes, short filter_indices[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_CLASSES]) { int tmp_clip[MAX_NUM_ALF_LUMA_COEFF]; int best_merge_clip[MAX_NUM_ALF_LUMA_COEFF]; double err[MAX_NUM_ALF_CLASSES]; double best_merge_err = MAX_DOUBLE; bool available_class[MAX_NUM_ALF_CLASSES]; int8_t index_list[MAX_NUM_ALF_CLASSES]; int8_t index_list_temp[MAX_NUM_ALF_CLASSES]; int num_remaining = num_classes; memset(filter_indices, 0, sizeof(short) * MAX_NUM_ALF_CLASSES * MAX_NUM_ALF_CLASSES); for (int i = 0; i < num_classes; i++) { filter_indices[num_remaining - 1][i] = i; index_list[i] = i; available_class[i] = true; copy_cov(&cov_merged[i], &cov[i]); cov_merged[i].num_bins = alf_aps->non_linear_flag[CHANNEL_TYPE_LUMA] ? MAX_ALF_NUM_CLIPPING_VALUES : 1; } // Try merging different covariance matrices // temporal AlfCovariance structure is allocated as the last element in covMerged array, the size of covMerged is MAX_NUM_ALF_CLASSES + 1 alf_covariance* tmp_cov = &cov_merged[MAX_NUM_ALF_CLASSES]; tmp_cov->num_bins = alf_aps->non_linear_flag[CHANNEL_TYPE_LUMA] ? MAX_ALF_NUM_CLIPPING_VALUES : 1; // init Clip for (int i = 0; i < num_classes; i++) { for (int val = 0; val < MAX_NUM_ALF_LUMA_COEFF; val++) { clip_merged[num_remaining - 1][i][val] = alf_aps->non_linear_flag[CHANNEL_TYPE_LUMA] ? MAX_ALF_NUM_CLIPPING_VALUES / 2 : 0; } if (alf_aps->non_linear_flag[CHANNEL_TYPE_LUMA]) { err[i] = optimize_filter_clip(&cov_merged[i], clip_merged[num_remaining - 1][i]); } else { err[i] = calculate_error_opt_filt(&cov_merged[i], clip_merged[num_remaining - 1][i]); } } while (num_remaining > 2) { double error_min = MAX_DOUBLE; //std::numeric_limits::max(); int best_to_merge_idx1 = 0, best_to_merge_idx2 = 1; for (int i = 0; i < num_classes - 1; i++) { if (available_class[i]) { for (int j = i + 1; j < num_classes; j++) { if (available_class[j]) { double error1 = err[i]; double error2 = err[j]; add_alf_cov_lhs_rhs(tmp_cov, &cov_merged[i], &cov_merged[j]); for (int l = 0; l < MAX_NUM_ALF_LUMA_COEFF; ++l) { tmp_clip[l] = (clip_merged[num_remaining - 1][i][l] + clip_merged[num_remaining - 1][j][l] + 1) >> 1; } double error_merged = alf_aps->non_linear_flag[CHANNEL_TYPE_LUMA] ? optimize_filter_clip(tmp_cov, tmp_clip) : calculate_error_opt_filt(tmp_cov, tmp_clip); double error = error_merged - error1 - error2; if (error < error_min) { best_merge_err = error_merged; memcpy(best_merge_clip, tmp_clip, sizeof(best_merge_clip)); error_min = error; best_to_merge_idx1 = i; best_to_merge_idx2 = j; } } } } } add_alf_cov(&cov_merged[best_to_merge_idx1], &cov_merged[best_to_merge_idx2]); memcpy(clip_merged[num_remaining - 2], clip_merged[num_remaining - 1], sizeof(int[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF])); memcpy(clip_merged[num_remaining - 2][best_to_merge_idx1], best_merge_clip, sizeof(best_merge_clip)); err[best_to_merge_idx1] = best_merge_err; available_class[best_to_merge_idx2] = false; for (int i = 0; i < num_classes; i++) { if (index_list[i] == best_to_merge_idx2) { index_list[i] = best_to_merge_idx1; } } num_remaining--; if (num_remaining <= num_classes) { memcpy(index_list_temp, index_list, sizeof(int8_t) * num_classes); bool exist = false; int ind = 0; for (int j = 0; j < num_classes; j++) { exist = false; for (int i = 0; i < num_classes; i++) { if (index_list_temp[i] == j) { exist = true; break; } } if (exist) { for (int i = 0; i < num_classes; i++) { if (index_list_temp[i] == j) { filter_indices[num_remaining - 1][i] = ind; index_list_temp[i] = -1; } } ind++; } } } } } static double alf_derive_coeff_quant(channel_type channel, int *filter_clipp, int *filter_coeff_quant, const alf_covariance* cov, const int bit_depth, const bool optimize_clip) { const bool is_luma = channel == CHANNEL_TYPE_LUMA ? true : false; const int num_coeff = is_luma ? 13 : 7; const int factor = 1 << (bit_depth - 1); const int max_value = factor - 1; const int min_value = -factor + 1; double filter_coeff[MAX_NUM_ALF_LUMA_COEFF]; optimize_filter(cov, filter_clipp, filter_coeff, optimize_clip); //roundFiltCoeff(filter_coeff_quant, filter_coeff, num_coeff, factor); for (int i = 0; i < num_coeff; i++) { int sign = filter_coeff[i] > 0 ? 1 : -1; filter_coeff_quant[i] = (int)(filter_coeff[i] * sign * factor + 0.5) * sign; } for (int i = 0; i < num_coeff - 1; i++) { filter_coeff_quant[i] = MIN(max_value, MAX(min_value, filter_coeff_quant[i])); } filter_coeff_quant[num_coeff - 1] = 0; int modified = 1; double err_ref = calc_error_for_coeffs(cov, filter_clipp, filter_coeff_quant, num_coeff, bit_depth); int sign; while (modified) { modified = 0; for (int sign_count = 0; sign_count <= 1; sign_count++) { sign = sign_count == 0 ? 1 : -1; double err_min = MAX_DOUBLE; int min_ind = -1; for (int k = 0; k < num_coeff - 1; k++) { if (filter_coeff_quant[k] - sign > max_value || filter_coeff_quant[k] - sign < min_value) { continue; } filter_coeff_quant[k] -= sign; double error = calc_error_for_coeffs(cov, filter_clipp, filter_coeff_quant, num_coeff, bit_depth); if (error < err_min) { err_min = error; min_ind = k; } filter_coeff_quant[k] += sign; } if (err_min < err_ref) { filter_coeff_quant[min_ind] -= sign; modified++; err_ref = err_min; } } } return err_ref; } static double alf_derive_filter_coeffs(alf_aps *aps, channel_type channel, alf_covariance *cov, alf_covariance *covMerged, short* filter_indices, int num_filters, double error_tab_force_0_coeff[MAX_NUM_ALF_CLASSES][2], int clip_merged[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF], int filter_coeff_set[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF], int filter_clipp_set[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF], const int bit_depth) { int num_coeff = channel == CHANNEL_TYPE_LUMA ? 13 : 7; double error = 0.0; alf_covariance *tmp_cov = &covMerged[MAX_NUM_ALF_CLASSES]; for (int filt_idx = 0; filt_idx < num_filters; filt_idx++) { reset_alf_covariance(tmp_cov, -1); bool found_clip = false; for (int class_idx = 0; class_idx < MAX_NUM_ALF_CLASSES; class_idx++) { if (filter_indices[class_idx] == filt_idx) { add_alf_cov(tmp_cov, &cov[class_idx]); if (!found_clip) { found_clip = true; // clip should be at the adress of shortest one memcpy(filter_clipp_set[filt_idx], clip_merged[num_filters - 1][class_idx], sizeof(int[MAX_NUM_ALF_LUMA_COEFF])); } } } // Find coeffcients assert(num_coeff == tmp_cov->num_coeff); error_tab_force_0_coeff[filt_idx][1] = tmp_cov->pix_acc + alf_derive_coeff_quant(channel, filter_clipp_set[filt_idx], filter_coeff_set[filt_idx], tmp_cov, bit_depth, false); error_tab_force_0_coeff[filt_idx][0] = tmp_cov->pix_acc; error += error_tab_force_0_coeff[filt_idx][1]; } return error; } static int alf_derive_filter_coefficients_prediction_mode(const alf_aps *alf_param, channel_type channel, const int num_filters, int filter_set[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF], int filter_clipp_set[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF]) { return (alf_param->non_linear_flag[CHANNEL_TYPE_LUMA] ? get_cost_filter_clipp(channel, num_filters, filter_set, filter_clipp_set) : 0) + get_cost_filter_coeff(channel, num_filters, filter_set); } static double alf_merge_filters_and_cost(encoder_state_t * const state, alf_aps *alf_aps, channel_type channel, int *ui_coeff_bits, alf_covariance *cov_frame, alf_covariance *cov_merged, int clip_merged[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF], array_variables *arr_vars) { const int num_coeff = channel == CHANNEL_TYPE_LUMA ? 13 : 7; int num_filters_best = 0; int num_filters = MAX_NUM_ALF_CLASSES; bool coded_var_bins[MAX_NUM_ALF_CLASSES]; double error_force_0_coeff_tab[MAX_NUM_ALF_CLASSES][2]; double lambda = state->frame->lambda; const int8_t bit_depth = state->encoder_control->bitdepth; double cost, cost0, dist, dist_force0, cost_min = MAX_DOUBLE; int coeff_bits, coeff_bits_force0; //clip_merged:iä ei tarvitse nollata ennen alf_merge_classes(alf_aps, channel, cov_frame, cov_merged, clip_merged, MAX_NUM_ALF_CLASSES, arr_vars->filter_indices); while (num_filters >= 1) { dist = alf_derive_filter_coeffs(alf_aps, channel, cov_frame, cov_merged, arr_vars->filter_indices[num_filters - 1], num_filters, error_force_0_coeff_tab, clip_merged, arr_vars->filter_coeff_set, arr_vars->filter_clipp_set, bit_depth); // filter coeffs are stored in m_filterCoeffSet dist_force0 = get_dist_force_0(alf_aps, channel, num_filters, error_force_0_coeff_tab, coded_var_bins, lambda, arr_vars->filter_coeff_set, arr_vars->filter_clipp_set); coeff_bits = alf_derive_filter_coefficients_prediction_mode(alf_aps, channel, num_filters, arr_vars->filter_coeff_set, arr_vars->filter_clipp_set); coeff_bits_force0 = get_cost_filter_coeff_force_0(alf_aps, channel, num_filters, coded_var_bins, arr_vars->filter_coeff_set, arr_vars->filter_clipp_set); cost = dist + lambda * coeff_bits; cost0 = dist_force0 + lambda * coeff_bits_force0; if (cost0 < cost) { cost = cost0; } if (cost <= cost_min) { cost_min = cost; num_filters_best = num_filters; } num_filters--; } dist = alf_derive_filter_coeffs(alf_aps, channel, cov_frame, cov_merged, arr_vars->filter_indices[num_filters_best - 1], num_filters_best, error_force_0_coeff_tab, clip_merged, arr_vars->filter_coeff_set, arr_vars->filter_clipp_set, bit_depth); coeff_bits = alf_derive_filter_coefficients_prediction_mode(alf_aps, channel, num_filters_best, arr_vars->filter_coeff_set, arr_vars->filter_clipp_set); dist_force0 = get_dist_force_0(alf_aps, channel, num_filters_best, error_force_0_coeff_tab, coded_var_bins, lambda, arr_vars->filter_coeff_set, arr_vars->filter_clipp_set); coeff_bits_force0 = get_cost_filter_coeff_force_0(alf_aps, channel, num_filters_best, coded_var_bins, arr_vars->filter_coeff_set, arr_vars->filter_clipp_set); cost = dist + lambda * coeff_bits; cost0 = dist_force0 + lambda * coeff_bits_force0; alf_aps->num_luma_filters = num_filters_best; double dist_return; if (cost <= cost0) { dist_return = dist; alf_aps->alf_luma_coeff_delta_flag = 0; *ui_coeff_bits = coeff_bits; } else { dist_return = dist_force0; alf_aps->alf_luma_coeff_delta_flag = 1; *ui_coeff_bits = coeff_bits_force0; memcpy(alf_aps->alf_luma_coeff_flag, coded_var_bins, sizeof(coded_var_bins)); for (int var_ind = 0; var_ind < num_filters_best; var_ind++) { if (coded_var_bins[var_ind] == 0) { memset(arr_vars->filter_coeff_set[var_ind], 0, sizeof(int) * MAX_NUM_ALF_LUMA_COEFF); memset(arr_vars->filter_clipp_set[var_ind], 0, sizeof(int) * MAX_NUM_ALF_LUMA_COEFF); } } } for (int ind = 0; ind < alf_aps->num_luma_filters; ++ind) { for (int i = 0; i < num_coeff; i++) { alf_aps->luma_coeff[ind * MAX_NUM_ALF_LUMA_COEFF + i] = arr_vars->filter_coeff_set[ind][i]; alf_aps->luma_clipp[ind * MAX_NUM_ALF_LUMA_COEFF + i] = arr_vars->filter_clipp_set[ind][i]; } } memcpy(alf_aps->filter_coeff_delta_idx, arr_vars->filter_indices[num_filters_best - 1], sizeof(short) * MAX_NUM_ALF_CLASSES); *ui_coeff_bits += get_non_filter_coeff_rate(alf_aps); return dist_return; } static double alf_get_filter_coeff_and_cost(encoder_state_t * const state, channel_type channel, double dist_unfilter, int *ui_coeff_bits, bool b_re_collect_stat, bool only_filter_cost, array_variables *arr_vars) { alf_info_t *alf_info = state->tile->frame->alf_info; alf_aps *alf_param_temp = &alf_info->alf_param_temp; cabac_data_t *cabac_estimator = &alf_info->cabac_estimator; bool is_luma = channel == CHANNEL_TYPE_LUMA ? 1 : 0; const int num_coeff = channel == CHANNEL_TYPE_LUMA ? 13 : 7; double lambda = state->frame->lambda; int clip_merged[MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_CLASSES][MAX_NUM_ALF_LUMA_COEFF]; const int8_t bit_depth = state->encoder_control->bitdepth; const int32_t num_ctus_in_pic = state->tile->frame->width_in_lcu * state->tile->frame->height_in_lcu; alf_covariance *alf_cov_frame = is_luma ? alf_info->alf_covariance_frame_luma : alf_info->alf_covariance_frame_chroma; //collect stat based on CTU decision if (b_re_collect_stat) { get_frame_stats(state->tile->frame->alf_info, channel, num_ctus_in_pic); } double dist = dist_unfilter; (*ui_coeff_bits) = 0; //get filter coeff if (is_luma) { const int fill_val = alf_param_temp->non_linear_flag[channel] ? MAX_ALF_NUM_CLIPPING_VALUES / 2 : 0; for (int i = 0; i < MAX_NUM_ALF_CLASSES; i++) { for (int j = 0; j < MAX_NUM_ALF_CLASSES; j++) { for (int k = 0; k < MAX_NUM_ALF_LUMA_COEFF; k++) { clip_merged[i][j][k] = fill_val; } } } // Reset Merge Tmp Cov reset_alf_covariance(&alf_info->alf_covariance_merged[MAX_NUM_ALF_CLASSES], MAX_ALF_NUM_CLIPPING_VALUES); reset_alf_covariance(&alf_info->alf_covariance_merged[MAX_NUM_ALF_CLASSES + 1], MAX_ALF_NUM_CLIPPING_VALUES); //distortion dist += alf_merge_filters_and_cost(state, alf_param_temp, channel, ui_coeff_bits, alf_cov_frame, alf_info->alf_covariance_merged, clip_merged, arr_vars); } else { //distortion for (int alt_idx = 0; alt_idx < alf_param_temp->num_alternatives_chroma; ++alt_idx) { assert(num_coeff == alf_cov_frame[alt_idx].num_coeff); alf_aps best_slice_param; double best_cost = MAX_DOUBLE; double best_dist = MAX_DOUBLE; int best_coeff_bits = 0; const int non_linear_flag_max = state->encoder_control->cfg.alf_non_linear_chroma ? 2 : 1; for (int non_linear_flag = 0; non_linear_flag < non_linear_flag_max; non_linear_flag++) { int current_non_linear_flag = alf_param_temp->non_linear_flag[channel] ? 1 : 0; if (non_linear_flag != current_non_linear_flag) { continue; } int fill_val = non_linear_flag ? MAX_ALF_NUM_CLIPPING_VALUES / 2 : 0; for (int i = 0; i < MAX_NUM_ALF_CHROMA_COEFF; i++) { arr_vars->filter_clipp_set[alt_idx][i] = fill_val; } double dist = alf_cov_frame[alt_idx].pix_acc + alf_derive_coeff_quant(channel, arr_vars->filter_clipp_set[alt_idx], arr_vars->filter_coeff_set[alt_idx], &alf_cov_frame[alt_idx], bit_depth, non_linear_flag); for (int i = 0; i < MAX_NUM_ALF_CHROMA_COEFF; i++) { alf_param_temp->chroma_coeff[alt_idx][i] = arr_vars->filter_coeff_set[alt_idx][i]; alf_param_temp->chroma_clipp[alt_idx][i] = arr_vars->filter_clipp_set[alt_idx][i]; } int coeff_bits = get_chroma_coeff_rate(alf_param_temp, alt_idx); double cost = dist + lambda * coeff_bits; if (cost < best_cost) { best_cost = cost; best_dist = dist; best_coeff_bits = coeff_bits; copy_alf_param(&best_slice_param, alf_param_temp); } } *ui_coeff_bits += best_coeff_bits; dist += best_dist; copy_alf_param(alf_param_temp, &best_slice_param); } (*ui_coeff_bits) += length_uvlc(alf_param_temp->num_alternatives_chroma - 1); (*ui_coeff_bits)++; } if (only_filter_cost) { return dist + lambda * (*ui_coeff_bits); } double rate = (*ui_coeff_bits); //m_CABACEstimator->resetBits(); alf_cabac_reset_bits(cabac_estimator); //m_CABACEstimator->codeAlfCtuEnableFlags(cs, channel, &m_alfParamTemp); code_alf_ctu_enable_flags_channel(state, cabac_estimator, channel, alf_param_temp); for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { if (is_luma) { // Evaluate cost of signaling filter set index for convergence of filters enabled flag / filter derivation assert(alf_info->alf_ctb_filter_index[ctu_idx] == ALF_NUM_FIXED_FILTER_SETS); assert(state->slice->alf->tile_group_num_aps == 1); //m_CABACEstimator->codeAlfCtuFilterIndex(cs, ctu_idx, &m_alfParamTemp.enabledFlag[COMPONENT_Y]); code_alf_ctu_filter_index(state, cabac_estimator, ctu_idx, alf_param_temp->enabled_flag[COMPONENT_Y]); } } //m_CABACEstimator->codeAlfCtuAlternatives(cs, channel, &m_alfParamTemp); code_alf_ctu_alternatives_channel(state, cabac_estimator, channel, alf_param_temp); rate += (23 - cabac_estimator->bits_left) + (cabac_estimator->num_buffered_bytes << 3); //frac_bits_scale * 0;/*(double)m_CABACEstimator->getEstFracBits();*/ return dist + lambda * rate; } static void alf_encoder(encoder_state_t * const state, alf_aps *aps, channel_type channel, const double lambda_chroma_weight, // = 0.0 array_variables *arr_vars) { alf_info_t *alf_info = state->tile->frame->alf_info; bool **ctu_enable_flags = alf_info->ctu_enable_flag; bool **ctu_enable_flags_tmp = alf_info->ctu_enable_flag_tmp; uint8_t **ctu_alternatives = alf_info->ctu_alternative; uint8_t **ctu_alternatives_tmp = alf_info->ctu_alternative_tmp; alf_aps *alf_param_temp = &alf_info->alf_param_temp; cabac_data_t ctx_start; cabac_data_t *cabac_estimator = &alf_info->cabac_estimator; memcpy(&ctx_start, cabac_estimator, sizeof(ctx_start)); //TempCtx ctxBest(m_CtxCache); cabac_data_t ctx_best; memcpy(&ctx_best, &ctx_start, sizeof(ctx_best)); bool is_luma = channel == CHANNEL_TYPE_LUMA ? 1 : 0; alf_covariance *alf_cov_frame = is_luma ? alf_info->alf_covariance_frame_luma : alf_info->alf_covariance_frame_chroma; kvz_config cfg = state->encoder_control->cfg; double cost_min = MAX_DOUBLE; double lambda = state->frame->lambda; unsigned *bits_new_filter = arr_vars->bits_new_filter; bits_new_filter[channel] = 0; const int num_classes = is_luma ? MAX_NUM_ALF_CLASSES : 1; const int32_t num_ctus_in_pic = state->tile->frame->width_in_lcu * state->tile->frame->height_in_lcu; int ui_coeff_bits = 0; //m_alfSliceParamTemp = alfSliceParam; copy_alf_param(alf_param_temp, aps); //1. get unfiltered distortion if (!is_luma) { alf_param_temp->num_alternatives_chroma = 1; } double cost = get_unfiltered_distortion_cov_channel(alf_cov_frame, channel); cost /= 1.001; // slight preference for unfiltered choice if (cost < cost_min) { cost_min = cost; if (is_luma) { aps->enabled_flag[COMPONENT_Y] = 0; } else { aps->enabled_flag[COMPONENT_Cb] = 0; aps->enabled_flag[COMPONENT_Cr] = 0; } // no CABAC signalling //ctxBest = AlfCtx(ctxStart); memcpy(&ctx_best, &ctx_start, sizeof(ctx_best)); //setCtuEnableFlag(m_ctuEnableFlagTmp, channel, 0); set_ctu_enable_flag(ctu_enable_flags_tmp, channel, 0, num_ctus_in_pic); if (!is_luma) { for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { ctu_alternatives_tmp[COMPONENT_Cb][ctu_idx] = 0; ctu_alternatives_tmp[COMPONENT_Cr][ctu_idx] = 0; } } } const int non_linear_flag_max = (is_luma ? cfg.alf_non_linear_luma : cfg.alf_non_linear_chroma) // For Chroma non linear flag is check for each alternative filter ? 2 : 1; for (int non_linear_flag = 0; non_linear_flag < non_linear_flag_max; non_linear_flag++) { for (int num_alternatives = is_luma ? 1 : MIN(num_ctus_in_pic * 2, MAX_NUM_ALF_ALTERNATIVES_CHROMA); num_alternatives > 0; num_alternatives--) { if (!is_luma) { alf_param_temp->num_alternatives_chroma = num_alternatives; } //2. all CTUs are on if (is_luma) { alf_param_temp->enabled_flag[COMPONENT_Y] = 1; } else { alf_param_temp->enabled_flag[COMPONENT_Cb] = 1; alf_param_temp->enabled_flag[COMPONENT_Cr] = 1; } alf_param_temp->non_linear_flag[channel] = non_linear_flag; //m_CABACEstimator->getCtx() = AlfCtx(ctxStart); memcpy(cabac_estimator, &ctx_start, sizeof(*cabac_estimator)); //setCtuEnableFlag(m_ctuEnableFlag, channel, 1); set_ctu_enable_flag(ctu_enable_flags, channel, 1, num_ctus_in_pic); // all alternatives are on if (!is_luma) { init_ctu_alternative_chroma(alf_param_temp, ctu_alternatives, num_ctus_in_pic); } cost = alf_get_filter_coeff_and_cost(state, channel, 0, &ui_coeff_bits, true, false, arr_vars); if (cost < cost_min) { bits_new_filter[channel] = ui_coeff_bits; cost_min = cost; copy_alf_param_w_channel(aps, alf_param_temp, channel); //ctxBest = AlfCtx(m_CABACEstimator->getCtx()); memcpy(&ctx_best, cabac_estimator, sizeof(ctx_best)); //setCtuEnableFlag(m_ctuEnableFlagTmp, channel, 1); set_ctu_enable_flag(ctu_enable_flags_tmp, channel, 1, num_ctus_in_pic); if (!is_luma) { memcpy(ctu_alternatives_tmp[COMPONENT_Cb], ctu_alternatives[COMPONENT_Cb], sizeof(uint8_t) * num_ctus_in_pic); memcpy(ctu_alternatives_tmp[COMPONENT_Cr], ctu_alternatives[COMPONENT_Cr], sizeof(uint8_t) * num_ctus_in_pic); } } //3. CTU decision double dist_unfilter = 0; double prev_it_cost = MAX_DOUBLE; const int iter_num = is_luma ? (2 * 4 + 1) : (2 * (2 + alf_param_temp->num_alternatives_chroma - 1) + 1); for (int iter = 0; iter < iter_num; iter++) { if ((iter & 0x01) == 0) { //m_CABACEstimator->getCtx() = AlfCtx(ctxStart); memcpy(cabac_estimator, &ctx_start, sizeof(*cabac_estimator)); cost = lambda * ui_coeff_bits; cost += alf_derive_ctb_alf_enable_flags(state, channel, &dist_unfilter, num_classes, lambda_chroma_weight, arr_vars); if (cost < cost_min) { bits_new_filter[channel] = ui_coeff_bits; cost_min = cost; //ctxBest = AlfCtx(m_CABACEstimator->getCtx()); memcpy(&ctx_best, cabac_estimator, sizeof(ctx_best)); //copyCtuEnableFlag(m_ctuEnableFlagTmp, m_ctuEnableFlag, channel); copy_ctu_enable_flag(ctu_enable_flags_tmp, ctu_enable_flags, channel, num_ctus_in_pic); if (!is_luma) { for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { ctu_alternatives_tmp[COMPONENT_Cb][ctu_idx] = ctu_alternatives[COMPONENT_Cb][ctu_idx]; ctu_alternatives_tmp[COMPONENT_Cr][ctu_idx] = ctu_alternatives[COMPONENT_Cr][ctu_idx]; } } copy_alf_param_w_channel(aps, alf_param_temp, channel); } else if (cost >= prev_it_cost) { // High probability that we have converged or we are diverging break; } prev_it_cost = cost; } else { // unfiltered distortion is added due to some CTBs may not use filter // no need to reset CABAC here, since uiCoeffBits is not affected /*cost = */alf_get_filter_coeff_and_cost(state, channel, dist_unfilter, &ui_coeff_bits, true, false, arr_vars); } }//for iter // Decrease number of alternatives and reset ctu params and filters } }//for non_linea_flag memcpy(cabac_estimator, &ctx_best, sizeof(*cabac_estimator)); if (!is_luma) { memcpy(ctu_alternatives[COMPONENT_Cb], ctu_alternatives_tmp[COMPONENT_Cb], sizeof(uint8_t) * num_ctus_in_pic); memcpy(ctu_alternatives[COMPONENT_Cr], ctu_alternatives_tmp[COMPONENT_Cr], sizeof(uint8_t) * num_ctus_in_pic); } copy_ctu_enable_flag(ctu_enable_flags, ctu_enable_flags_tmp, channel, num_ctus_in_pic); } static void alf_get_avai_aps_ids_luma(encoder_state_t * const state, int *new_aps_id, int *aps_ids, int *size_of_aps_ids, short alf_clipping_values[MAX_NUM_CHANNEL_TYPE][MAX_ALF_NUM_CLIPPING_VALUES]) { //alf_aps *apss = state->slice->alf->apss; for (int i = 0; i < ALF_CTB_MAX_NUM_APS; i++) { param_set_map* param_set = &state->tile->frame->alf_param_set_map[i + NUM_APS_TYPE_LEN + T_ALF_APS]; if (param_set->b_changed && (param_set->parameter_set.aps_id >= 0 || param_set->parameter_set.aps_id < ALF_CTB_MAX_NUM_APS)) { copy_aps(&state->slice->alf->apss[i], ¶m_set->parameter_set, false); } } //std::vector result; int aps_id_checked = 0, cur_aps_id = state->tile->frame->alf_info->aps_id_start; if (cur_aps_id < ALF_CTB_MAX_NUM_APS) { while ((aps_id_checked < ALF_CTB_MAX_NUM_APS) && !state->frame->is_irap && *size_of_aps_ids < ALF_CTB_MAX_NUM_APS /*&& /*!cs.slice->getPendingRasInit()*/) { alf_aps *cur_aps = &state->slice->alf->apss[cur_aps_id]; bool aps_found = (0 <= cur_aps->aps_id && cur_aps->aps_id < ALF_CTB_MAX_NUM_APS); if (aps_found/*cur_aps*/ && cur_aps->layer_id == 0/*cs.slice->getPic()->layerId*/ && cur_aps->temporal_id <= state->slice->id /*cs.slice->getTLayer(*/ && cur_aps->new_filter_flag[CHANNEL_TYPE_LUMA]) { for (int id = 0; id < ALF_CTB_MAX_NUM_APS; id++) { if (aps_ids[id] == -1) { aps_ids[id] = cur_aps_id; (*size_of_aps_ids)++; break; } } } aps_id_checked++; cur_aps_id = (cur_aps_id + 1) % ALF_CTB_MAX_NUM_APS; } } state->slice->alf->tile_group_num_aps = *size_of_aps_ids; for (int i = 0; i < state->slice->alf->tile_group_num_aps; i++) { state->slice->alf->tile_group_luma_aps_id[i] = aps_ids[i]; } //*new_aps_id = g_aps_id_start - 1; *new_aps_id = state->tile->frame->alf_info->aps_id_start - 1; if (*new_aps_id < 0) { *new_aps_id = (int)ALF_CTB_MAX_NUM_APS - 1; } assert(*new_aps_id < (int)ALF_CTB_MAX_NUM_APS); //Wrong APS index assignment in getAvaiApsIdsLuma } static void alf_derive_stats_for_filtering(encoder_state_t * const state, short alf_clipping_values[MAX_NUM_CHANNEL_TYPE][MAX_ALF_NUM_CLIPPING_VALUES]) { alf_info_t *alf_info = state->tile->frame->alf_info; enum kvz_chroma_format chroma_fmt = state->encoder_control->chroma_format; bool chroma_scale_x = (chroma_fmt == KVZ_CSP_444) ? 0 : 1; bool chroma_scale_y = (chroma_fmt != KVZ_CSP_420) ? 0 : 1; const int32_t num_ctus_in_pic = state->tile->frame->width_in_lcu * state->tile->frame->height_in_lcu; const int alf_vb_luma_ctu_height = LCU_WIDTH; const int alf_vb_chma_ctu_height = (LCU_WIDTH >> ((chroma_fmt == KVZ_CSP_420) ? 1 : 0)); const int alf_vb_luma_pos = LCU_WIDTH - ALF_VB_POS_ABOVE_CTUROW_LUMA; const int alf_vb_chma_pos = (LCU_WIDTH >> ((chroma_fmt == KVZ_CSP_420) ? 1 : 0)) - ALF_VB_POS_ABOVE_CTUROW_CHMA; int32_t pic_width = state->tile->frame->width; int32_t pic_height = state->tile->frame->height; int ctu_rs_addr = 0; const int number_of_components = (chroma_fmt == KVZ_CSP_400) ? 1 : MAX_NUM_COMPONENT; // init CTU stats buffers { for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { for (int class_idx = 0; class_idx < MAX_NUM_ALF_CLASSES; class_idx++) { reset_alf_covariance(&state->tile->frame->alf_info->alf_covariance_y[(ctu_idx * MAX_NUM_ALF_CLASSES) + class_idx], MAX_ALF_NUM_CLIPPING_VALUES); } } for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { reset_alf_covariance(&state->tile->frame->alf_info->alf_covariance_u[ctu_idx], MAX_ALF_NUM_CLIPPING_VALUES); reset_alf_covariance(&state->tile->frame->alf_info->alf_covariance_v[ctu_idx], MAX_ALF_NUM_CLIPPING_VALUES); } } // init Frame stats buffers const int number_of_channels = (chroma_fmt == KVZ_CSP_400) ? 1 : MAX_NUM_CHANNEL_TYPE; for (int channel_idx = 0; channel_idx < number_of_channels; channel_idx++) { const channel_type channel_id = channel_idx; const int num_classes = channel_id == CHANNEL_TYPE_LUMA ? MAX_NUM_ALF_CLASSES : 1; for (int class_idx = 0; class_idx < num_classes; class_idx++) { reset_alf_covariance(&alf_info->alf_covariance_frame_luma[class_idx], MAX_ALF_NUM_CLIPPING_VALUES); } reset_alf_covariance(&alf_info->alf_covariance_frame_chroma[0], MAX_ALF_NUM_CLIPPING_VALUES); } alf_covariance* alf_cov; alf_covariance* alf_cov_frame; for (int y_pos = 0; y_pos < pic_height; y_pos += LCU_WIDTH) { for (int x_pos = 0; x_pos < pic_width; x_pos += LCU_WIDTH) { const int width = (x_pos + LCU_WIDTH > pic_width) ? (pic_width - x_pos) : LCU_WIDTH; const int height = (y_pos + LCU_WIDTH > pic_height) ? (pic_height - y_pos) : LCU_WIDTH; for (int comp_idx = 0; comp_idx < number_of_components; comp_idx++) { alf_cov = comp_idx == COMPONENT_Y ? state->tile->frame->alf_info->alf_covariance_y : comp_idx == COMPONENT_Cb ? state->tile->frame->alf_info->alf_covariance_u : comp_idx == COMPONENT_Cr ? state->tile->frame->alf_info->alf_covariance_v : NULL; if (alf_cov == NULL) { assert(0); } const bool is_luma = comp_idx == COMPONENT_Y ? 1 : 0; channel_type ch_type = is_luma ? CHANNEL_TYPE_LUMA : CHANNEL_TYPE_CHROMA; alf_cov_frame = is_luma ? alf_info->alf_covariance_frame_luma : alf_info->alf_covariance_frame_chroma; int blk_w = is_luma ? width : width >> chroma_scale_x; int blk_h = is_luma ? height : height >> chroma_scale_y; int pos_x = is_luma ? x_pos : x_pos >> chroma_scale_x; int pos_y = is_luma ? y_pos : y_pos >> chroma_scale_y; int32_t org_stride = is_luma ? state->tile->frame->source->stride : state->tile->frame->source->stride >> chroma_scale_x; int32_t rec_stride = is_luma ? state->tile->frame->rec->stride : state->tile->frame->rec->stride >> chroma_scale_x; kvz_pixel *org = comp_idx ? (comp_idx - 1 ? &state->tile->frame->source->v[pos_x + pos_y * org_stride] : &state->tile->frame->source->u[pos_x + pos_y * org_stride]) : &state->tile->frame->source->y[pos_x + pos_y * org_stride]; kvz_pixel *rec = comp_idx ? (comp_idx - 1 ? &state->tile->frame->rec->v[pos_x + pos_y * rec_stride] : &state->tile->frame->rec->u[pos_x + pos_y * rec_stride]) : &state->tile->frame->rec->y[pos_x + pos_y * rec_stride]; const int num_classes = is_luma ? MAX_NUM_ALF_CLASSES : 1; const int cov_index = ctu_rs_addr * num_classes; kvz_alf_get_blk_stats(state, ch_type, &alf_cov[cov_index], comp_idx ? NULL : alf_info->classifier, org, org_stride, rec, rec_stride, pos_x, pos_y, pos_x, pos_y, blk_w, blk_h, (is_luma ? alf_vb_luma_ctu_height : alf_vb_chma_ctu_height), (is_luma) ? alf_vb_luma_pos : alf_vb_chma_pos, alf_clipping_values ); for (int class_idx = 0; class_idx < num_classes; class_idx++) { add_alf_cov(&alf_cov_frame[is_luma ? class_idx : 0], &alf_cov[cov_index + class_idx] ); } } ctu_rs_addr++; } } } static void alf_reconstruct_coeff_aps(encoder_state_t * const state, bool luma, bool chroma, bool is_rdo, array_variables *arr_vars) { //luma alf_aps* apss = state->slice->alf->apss; //AlfSliceParam alfSliceParamTmp; alf_aps alf_param_tmp; //APS* cur_aps; alf_aps* cur_aps; if (luma) { for (int i = 0; i < state->slice->alf->tile_group_num_aps /* 1, cs.slice->getTileGroupNumAps()*/; i++) { int aps_idx = state->slice->alf->tile_group_luma_aps_id[i]; cur_aps = &apss[aps_idx]; assert(cur_aps != NULL); // "invalid APS" alf_param_tmp = *cur_aps; alf_reconstruct_coeff(state, &alf_param_tmp, CHANNEL_TYPE_LUMA, is_rdo, true, arr_vars); memcpy(arr_vars->coeff_aps_luma[i], arr_vars->coeff_final, sizeof(arr_vars->coeff_final)); memcpy(arr_vars->clipp_aps_luma[i], arr_vars->clipp_final, sizeof(arr_vars->clipp_final)); } } //chroma if (chroma) { int aps_idx_chroma = state->slice->alf->tile_group_chroma_aps_id; cur_aps = &apss[aps_idx_chroma]; //copy_alf_param(g_alf_aps_chroma, cur_aps); //copy_alf_param(&alf_param_tmp, g_alf_aps_chroma); copy_alf_param(&alf_param_tmp, cur_aps); alf_reconstruct_coeff(state, &alf_param_tmp, CHANNEL_TYPE_CHROMA, is_rdo, true, arr_vars); } } static void alf_encoder_ctb(encoder_state_t * const state, alf_aps *aps, const double lambda_chroma_weight, array_variables *arr_vars) { alf_info_t *alf_info = state->tile->frame->alf_info; bool **ctu_enable_flag = alf_info->ctu_enable_flag; bool **ctu_enable_flag_tmp = alf_info->ctu_enable_flag_tmp; uint8_t **ctu_alternatives = alf_info->ctu_alternative; uint8_t **ctu_alternatives_tmp = alf_info->ctu_alternative_tmp; double **ctb_distortions_unfilter = alf_info->ctb_distortion_unfilter; unsigned *arr_bits_new_filter = arr_vars->bits_new_filter; short *alf_ctb_filter_index = alf_info->alf_ctb_filter_index; short *alf_ctb_filter_set_index_tmp = alf_info->alf_ctb_filter_set_index_tmp; alf_aps *alf_param_temp = &alf_info->alf_param_temp; cabac_data_t ctx_start; cabac_data_t *cabac_estimator = &alf_info->cabac_estimator; memcpy(&ctx_start, cabac_estimator, sizeof(ctx_start)); cabac_data_t ctx_best; memcpy(&ctx_best, &ctx_start, sizeof(ctx_best)); cabac_data_t ctx_temp_start; cabac_data_t ctx_temp_best; memcpy(&ctx_temp_best, &ctx_start, sizeof(ctx_temp_best)); cabac_data_t ctx_temp_alt_start; //cabac_data_t ctx_temp_alt_best; int best_aps_ids[ALF_CTB_MAX_NUM_APS] = { -1, -1, -1, -1, -1, -1, -1, -1 }; int filter_tmp[MAX_NUM_ALF_LUMA_COEFF] = { 0 }; int g_clip_tmp[MAX_NUM_ALF_LUMA_COEFF] = { 0 }; int size_of_best_aps_ids = 0; int clip_default[MAX_NUM_ALF_LUMA_COEFF] = { 0 }; const int8_t bit_depth = state->encoder_control->bitdepth; double lambda = state->frame->lambda; int size_of_aps_ids = 0; const int32_t num_ctus_in_pic = state->tile->frame->width_in_lcu * state->tile->frame->height_in_lcu; alf_aps alf_aps_temp_nl; alf_covariance *alf_cov_chroma; int cov_indx = 0; //AlfSliceParam alfSliceParamNewFiltersBest = alfSliceParamNewFilters; alf_aps alf_aps_new_filters_best; copy_alf_param(&alf_aps_new_filters_best, aps); alf_aps* apss = state->slice->alf->apss; bool has_new_filters[2] = { aps->enabled_flag[COMPONENT_Y] , aps->enabled_flag[COMPONENT_Cb] || aps->enabled_flag[COMPONENT_Cr] }; //initDistortion(); for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { ctb_distortions_unfilter[COMPONENT_Y][ctu_idx] = get_unfiltered_distortion_cov_classes(&alf_info->alf_covariance_y[ctu_idx], MAX_NUM_ALF_CLASSES); ctb_distortions_unfilter[COMPONENT_Cb][ctu_idx] = get_unfiltered_distortion_cov_classes(&alf_info->alf_covariance_u[ctu_idx], 1); ctb_distortions_unfilter[COMPONENT_Cr][ctu_idx] = get_unfiltered_distortion_cov_classes(&alf_info->alf_covariance_v[ctu_idx], 1); } //luma copy_alf_param(alf_param_temp, aps); memset(ctu_enable_flag[COMPONENT_Y], 1, sizeof(bool) * num_ctus_in_pic); get_frame_stats(alf_info, CHANNEL_TYPE_LUMA, num_ctus_in_pic); memset(ctu_enable_flag[COMPONENT_Y], 0, sizeof(bool) * num_ctus_in_pic); double cost_off = get_unfiltered_distortion_cov_channel(alf_info->alf_covariance_frame_luma, CHANNEL_TYPE_LUMA); int new_aps_id; int aps_ids[ALF_CTB_MAX_NUM_APS]; for (int i = 0; i < ALF_CTB_MAX_NUM_APS; i++) { aps_ids[i] = -1; } alf_get_avai_aps_ids_luma(state, &new_aps_id, aps_ids, &size_of_aps_ids, arr_vars->alf_clipping_values); double cost_min = MAX_DOUBLE; alf_reconstruct_coeff_aps(state, true, false, true, arr_vars); int num_loops = has_new_filters[CHANNEL_TYPE_LUMA] ? 2 : 1; for (int use_new_filter = 0; use_new_filter < num_loops; use_new_filter++) { int bits_new_filter = 0; if (use_new_filter == 1) { if (!has_new_filters[CHANNEL_TYPE_LUMA]) { continue; } else { bits_new_filter = arr_bits_new_filter[CHANNEL_TYPE_LUMA]; alf_reconstruct_coeff(state, aps, CHANNEL_TYPE_LUMA, true, true, arr_vars); } } int num_iter = use_new_filter ? 2 : 1; for (int num_temporal_aps = 0; num_temporal_aps <= size_of_aps_ids/*apsIds.size()*/; num_temporal_aps++) { if (num_temporal_aps + use_new_filter >= ALF_CTB_MAX_NUM_APS) { continue; } //cs.slice->setTileGroupNumAps(numTemporalAps + useNewFilter); state->slice->alf->tile_group_num_aps = num_temporal_aps + use_new_filter; int num_filter_set = ALF_NUM_FIXED_FILTER_SETS + num_temporal_aps + use_new_filter; if (num_temporal_aps == size_of_aps_ids && num_temporal_aps > 0 && use_new_filter && new_aps_id == aps_ids[size_of_aps_ids - 1] /*apsIds.back()*/) //last temporalAPS is occupied by new filter set and this temporal APS becomes unavailable { continue; } for (int iter = 0; iter < num_iter; iter++) { copy_alf_param(alf_param_temp, aps); alf_param_temp->enabled_flag[CHANNEL_TYPE_LUMA] = true; double cur_cost = 3 * lambda; if (iter > 0) //re-derive new filter-set { double d_dist_org_new_filter = 0; int blocks_using_new_filter = 0; for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { if (ctu_enable_flag[COMPONENT_Y][ctu_idx] && alf_ctb_filter_index[ctu_idx] != ALF_NUM_FIXED_FILTER_SETS) { ctu_enable_flag[COMPONENT_Y][ctu_idx] = 0; } else if (ctu_enable_flag[COMPONENT_Y][ctu_idx] && alf_ctb_filter_index[ctu_idx] == ALF_NUM_FIXED_FILTER_SETS) { blocks_using_new_filter++; d_dist_org_new_filter += ctb_distortions_unfilter[COMPONENT_Y][ctu_idx]; cov_indx = ctu_idx * MAX_NUM_ALF_CLASSES; for (int class_idx = 0; class_idx < MAX_NUM_ALF_CLASSES; class_idx++) { short* p_coeff = arr_vars->coeff_final; int16_t* p_clipp = arr_vars->clipp_final; for (int i = 0; i < MAX_NUM_ALF_LUMA_COEFF; i++) { filter_tmp[i] = p_coeff[class_idx * MAX_NUM_ALF_LUMA_COEFF + i]; g_clip_tmp[i] = p_clipp[class_idx * MAX_NUM_ALF_LUMA_COEFF + i]; } d_dist_org_new_filter += calc_error_for_coeffs(&alf_info->alf_covariance_y[cov_indx + class_idx], g_clip_tmp, filter_tmp, MAX_NUM_ALF_LUMA_COEFF, bit_depth); } } } //for ctb if (blocks_using_new_filter > 0 && blocks_using_new_filter < num_ctus_in_pic) { int bit_nl[2] = { 0, 0 }; double err_nl[2] = { 0.0, 0.0 }; alf_param_temp->non_linear_flag[CHANNEL_TYPE_LUMA] = 1; if (state->encoder_control->cfg.alf_non_linear_luma) { err_nl[1] = alf_get_filter_coeff_and_cost(state, CHANNEL_TYPE_LUMA, 0, &bit_nl[1], true, true, arr_vars); copy_alf_param(&alf_aps_temp_nl, alf_param_temp); } else { err_nl[1] = MAX_DOUBLE; } alf_param_temp->non_linear_flag[CHANNEL_TYPE_LUMA] = 0; err_nl[0] = alf_get_filter_coeff_and_cost(state, CHANNEL_TYPE_LUMA, 0, &bit_nl[0], true, true, arr_vars); int bits_new_filter_temp_luma = bit_nl[0]; double err = err_nl[0]; if (err_nl[1] < err_nl[0]) { err = err_nl[1]; bits_new_filter_temp_luma = bit_nl[1]; copy_alf_param(alf_param_temp, &alf_aps_temp_nl); } if (d_dist_org_new_filter + lambda * arr_bits_new_filter[CHANNEL_TYPE_LUMA] < err) //re-derived filter is not good, skip { continue; } alf_reconstruct_coeff(state, alf_param_temp, CHANNEL_TYPE_LUMA, true, true, arr_vars); bits_new_filter = bits_new_filter_temp_luma; } else //no blocks using new filter, skip { continue; } } //m_CABACEstimator->getCtx() = ctxStart; memcpy(cabac_estimator, &ctx_start, sizeof(*cabac_estimator)); for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { double dist_unfilter_ctb = ctb_distortions_unfilter[COMPONENT_Y][ctu_idx]; //ctb on ctu_enable_flag[COMPONENT_Y][ctu_idx] = 1; double cost_on = MAX_DOUBLE; //ctxTempStart = AlfCtx(m_CABACEstimator->getCtx()); memcpy(&ctx_temp_start, cabac_estimator, sizeof(ctx_temp_start)); ctx_temp_start.only_count = 1; int i_best_filter_set_idx = 0; int first_filter_set_idx = 0; if (!state->encoder_control->cfg.alf_allow_predefined_filters) { first_filter_set_idx = ALF_NUM_FIXED_FILTER_SETS; } for (int filter_set_idx = first_filter_set_idx; filter_set_idx < num_filter_set; filter_set_idx++) { //rate //m_CABACEstimator->getCtx() = AlfCtx(ctxTempStart); memcpy(cabac_estimator, &ctx_temp_start, sizeof(*cabac_estimator)); //m_CABACEstimator->resetBits(); alf_cabac_reset_bits(cabac_estimator); //m_CABACEstimator->codeAlfCtuEnableFlag(cs, ctbIdx, COMPONENT_Y, &m_alfSliceParamTemp); code_alf_ctu_enable_flag(state, cabac_estimator, ctu_idx, COMPONENT_Y, alf_param_temp); alf_ctb_filter_index[ctu_idx] = filter_set_idx; code_alf_ctu_filter_index(state, cabac_estimator, ctu_idx, alf_param_temp->enabled_flag[COMPONENT_Y]); double rate_on = (23 - cabac_estimator->bits_left) + (cabac_estimator->num_buffered_bytes << 3); //frac_bits_scale * 0; /*(double)m_CABACEstimator->getEstFracBits()*/ ; //distortion double dist = dist_unfilter_ctb; cov_indx = ctu_idx * MAX_NUM_ALF_CLASSES; for (int class_idx = 0; class_idx < MAX_NUM_ALF_CLASSES; class_idx++) { if (filter_set_idx < ALF_NUM_FIXED_FILTER_SETS) { int filter_idx = g_class_to_filter_mapping[filter_set_idx][class_idx]; dist += calc_error_for_coeffs(&alf_info->alf_covariance_y[cov_indx + class_idx], clip_default, g_fixed_filter_set_coeff[filter_idx], MAX_NUM_ALF_LUMA_COEFF, bit_depth); } else { short *p_coeff; int16_t *p_clipp; if (use_new_filter && filter_set_idx == ALF_NUM_FIXED_FILTER_SETS) { p_coeff = arr_vars->coeff_final; p_clipp = arr_vars->clipp_final; } else if (use_new_filter) { p_coeff = arr_vars->coeff_aps_luma[filter_set_idx - 1 - ALF_NUM_FIXED_FILTER_SETS]; p_clipp = arr_vars->clipp_aps_luma[filter_set_idx - 1 - ALF_NUM_FIXED_FILTER_SETS]; } else { p_coeff = arr_vars->coeff_aps_luma[filter_set_idx - ALF_NUM_FIXED_FILTER_SETS]; p_clipp = arr_vars->clipp_aps_luma[filter_set_idx - ALF_NUM_FIXED_FILTER_SETS]; } for (int i = 0; i < MAX_NUM_ALF_LUMA_COEFF; i++) { filter_tmp[i] = p_coeff[class_idx * MAX_NUM_ALF_LUMA_COEFF + i]; g_clip_tmp[i] = p_clipp[class_idx * MAX_NUM_ALF_LUMA_COEFF + i]; } dist += calc_error_for_coeffs(&alf_info->alf_covariance_y[cov_indx + class_idx], g_clip_tmp, filter_tmp, MAX_NUM_ALF_LUMA_COEFF, bit_depth); } } //cost double cost_on_tmp = dist + lambda * rate_on; if (cost_on_tmp < cost_on) { //ctxTempBest = AlfCtx(m_CABACEstimator->getCtx()); memcpy(&ctx_temp_best, cabac_estimator, sizeof(ctx_temp_best)); ctx_temp_best.only_count = 1; cost_on = cost_on_tmp; i_best_filter_set_idx = filter_set_idx; } } //ctb off ctu_enable_flag[COMPONENT_Y][ctu_idx] = 0; //rate //m_CABACEstimator->getCtx() = AlfCtx(ctxTempStart); memcpy(cabac_estimator, &ctx_temp_start, sizeof(*cabac_estimator)); //m_CABACEstimator->resetBits(); alf_cabac_reset_bits(cabac_estimator); //m_CABACEstimator->codeAlfCtuEnableFlag(cs, ctbIdx, COMPONENT_Y, &m_alfSliceParamTemp); code_alf_ctu_enable_flag(state, cabac_estimator, ctu_idx, COMPONENT_Y, alf_param_temp); //cost double cost_off = dist_unfilter_ctb + lambda * (23 - cabac_estimator->bits_left) + (cabac_estimator->num_buffered_bytes << 3);// frac_bits_scale * 0; /* (double)m_CABACEstimator->getEstFracBits()*/ ; if (cost_on < cost_off) { //m_CABACEstimator->getCtx() = AlfCtx(ctxTempBest); memcpy(cabac_estimator, &ctx_temp_best, sizeof(*cabac_estimator)); ctu_enable_flag[COMPONENT_Y][ctu_idx] = 1; alf_ctb_filter_index[ctu_idx] = i_best_filter_set_idx; cur_cost += cost_on; } else { ctu_enable_flag[COMPONENT_Y][ctu_idx] = 0; cur_cost += cost_off; } } //for(ctbIdx) int tmp_bits = bits_new_filter + 3 * (num_filter_set - ALF_NUM_FIXED_FILTER_SETS); cur_cost += tmp_bits * lambda; if (cur_cost < cost_min) { cost_min = cur_cost; size_of_best_aps_ids = num_filter_set - ALF_NUM_FIXED_FILTER_SETS; for (int i = 0; i < size_of_best_aps_ids; i++) { if (i == 0 && use_new_filter) { best_aps_ids[i] = new_aps_id; } else { best_aps_ids[i] = aps_ids[i - use_new_filter]; } } //alfSliceParamNewFiltersBest = m_alfSliceParamTemp; copy_alf_param(&alf_aps_new_filters_best, alf_param_temp); //ctxBest = AlfCtx(m_CABACEstimator->getCtx()); memcpy(&ctx_best, cabac_estimator, sizeof(ctx_best)); //copyCtuEnableFlag(m_ctuEnableFlagTmp, m_ctuEnableFlag, CHANNEL_TYPE_LUMA); memcpy(ctu_enable_flag_tmp[COMPONENT_Y], ctu_enable_flag[COMPONENT_Y], sizeof(bool) * num_ctus_in_pic); for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { alf_ctb_filter_set_index_tmp[ctu_idx] = alf_ctb_filter_index[ctu_idx]; } alf_aps_new_filters_best.new_filter_flag[CHANNEL_TYPE_LUMA] = use_new_filter; } }//for (int iter = 0; iter < numIter; iter++) }// for (int numTemporalAps = 0; numTemporalAps < apsIds.size(); numTemporalAps++) }//for (int useNewFilter = 0; useNewFilter <= 1; useNewFilter++) state->slice->alf->tile_group_cc_alf_cb_aps_id = new_aps_id; state->slice->alf->tile_group_cc_alf_cr_aps_id = new_aps_id; if (cost_off <= cost_min) { memset(state->slice->alf->tile_group_alf_enabled_flag, 0, sizeof(state->slice->alf->tile_group_alf_enabled_flag)); state->slice->alf->tile_group_num_aps = 0; for (int i = 0; i < MAX_NUM_COMPONENT; i++) { memset(ctu_enable_flag[i], 0, sizeof(bool) * num_ctus_in_pic); } return; } else { //cs.slice->setTileGroupAlfEnabledFlag(COMPONENT_Y, true); state->slice->alf->tile_group_alf_enabled_flag[COMPONENT_Y] = true; //cs.slice->setTileGroupNumAps((int)bestApsIds.size()); state->slice->alf->tile_group_num_aps = size_of_best_aps_ids; //cs.slice->setAPSs(bestApsIds); for (int i = 0; i < size_of_best_aps_ids; i++) { state->slice->alf->tile_group_luma_aps_id[i] = best_aps_ids[i]; } //copyCtuEnableFlag(m_ctuEnableFlag, m_ctuEnableFlagTmp, CHANNEL_TYPE_LUMA); copy_ctu_enable_flag(ctu_enable_flag, ctu_enable_flag_tmp, CHANNEL_TYPE_LUMA, num_ctus_in_pic); for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { alf_ctb_filter_index[ctu_idx] = alf_ctb_filter_set_index_tmp[ctu_idx]; } if (alf_aps_new_filters_best.new_filter_flag[CHANNEL_TYPE_LUMA]) { //APS* newAPS = m_apsMap->getPS((new_aps_id << NUM_APS_TYPE_LEN) + T_ALF_APS); alf_aps* new_aps = &state->tile->frame->alf_param_set_map[new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set; if (new_aps->aps_id < 0 || new_aps->aps_id >= ALF_CTB_MAX_NUM_APS) // new_aps == NULL { //newAPS = m_apsMap->allocatePS(new_aps_id); assert(new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS < ALF_CTB_MAX_NUM_APS); //Invalid PS id bool found = false; for (int i = 0; i < ALF_CTB_MAX_NUM_APS; i++) { if (state->tile->frame->alf_param_set_map[i].parameter_set.aps_id == new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS) { found = true; } } if (!found) { state->tile->frame->alf_param_set_map[new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].b_changed = true; //state->tile->frame->alf_param_set_map[new_aps_id + NUM_APS_TYPE_LEN+ T_ALF_APS].p_nalu_data = 0; //state->tile->frame->alf_param_set_map[new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set = malloc(sizeof(alf_aps)); state->tile->frame->alf_param_set_map[new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set.aps_id = new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS; } //copy_alf_param(new_aps, &state->tile->frame->alf_param_set_map[new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set); new_aps->aps_id = new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS; new_aps->aps_type = T_ALF_APS; } copy_alf_param(new_aps, &alf_aps_new_filters_best); new_aps->temporal_id = state->slice->id; new_aps->new_filter_flag[CHANNEL_TYPE_CHROMA] = false; state->tile->frame->alf_param_set_map[new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].b_changed = true; alf_info->aps_id_start = new_aps_id; } int8_t *aps_ids = state->slice->alf->tile_group_luma_aps_id; for (int i = 0; i < state->slice->alf->tile_group_num_aps; i++) { copy_aps(&apss[aps_ids[i]], &state->tile->frame->alf_param_set_map[aps_ids[i] + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set, false); } } //chroma if (state->encoder_control->chroma_format != KVZ_CSP_400) { copy_alf_param(alf_param_temp, &alf_aps_new_filters_best); if (alf_param_temp->num_alternatives_chroma < 1) { alf_param_temp->num_alternatives_chroma = 1; } for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { ctu_alternatives[COMPONENT_Cb][ctu_idx] = 0; ctu_alternatives[COMPONENT_Cr][ctu_idx] = 0; } set_ctu_enable_flag(ctu_enable_flag, CHANNEL_TYPE_CHROMA, 1, num_ctus_in_pic); get_frame_stats(alf_info, CHANNEL_TYPE_CHROMA, num_ctus_in_pic); cost_off = get_unfiltered_distortion_cov_channel(alf_info->alf_covariance_frame_chroma, CHANNEL_TYPE_CHROMA); cost_min = MAX_DOUBLE; //m_CABACEstimator->getCtx() = AlfCtx(ctxBest); memcpy(cabac_estimator, &ctx_best, sizeof(*cabac_estimator)); //ctxStart = AlfCtx(m_CABACEstimator->getCtx()); memcpy(&ctx_start, cabac_estimator, sizeof(ctx_start)); ctx_start.only_count = 1; int new_aps_id_chroma = -1; if (alf_aps_new_filters_best.new_filter_flag[CHANNEL_TYPE_LUMA] && (alf_aps_new_filters_best.enabled_flag[COMPONENT_Cb] || alf_aps_new_filters_best.enabled_flag[COMPONENT_Cr])) { new_aps_id_chroma = new_aps_id; } else if (alf_aps_new_filters_best.enabled_flag[COMPONENT_Cb] || alf_aps_new_filters_best.enabled_flag[COMPONENT_Cr]) { int cur_id = alf_info->aps_id_start; if (size_of_aps_ids < 8 || state->slice->alf->tile_group_num_aps < 8) { while (new_aps_id_chroma < 0) { cur_id--; if (cur_id < 0) { cur_id = ALF_CTB_MAX_NUM_APS - 1; } bool found = false; for (int i = 0; i < 8; i++) { if (cur_id == best_aps_ids[i]) { found = true; } } if (!found) { new_aps_id_chroma = cur_id; } } } } for (int cur_aps_id = 0; cur_aps_id < ALF_CTB_MAX_NUM_APS; cur_aps_id++) { const bool reuse_existing_aps = cur_aps_id != new_aps_id_chroma; if ((/*(cs.slice->getPendingRasInit() ||*/ state->frame->is_irap) && reuse_existing_aps) { continue; } alf_aps* cur_aps = &state->tile->frame->alf_param_set_map[cur_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set; double cur_cost = lambda * 3; if (!reuse_existing_aps) { copy_alf_param(alf_param_temp, aps); cur_cost += lambda * arr_bits_new_filter[CHANNEL_TYPE_CHROMA]; } else if (cur_aps && cur_aps->temporal_id <= state->slice->id && cur_aps->layer_id != 0 && cur_aps->new_filter_flag[CHANNEL_TYPE_CHROMA]) { //g_alf_slice_aps_temp = cur_aps; copy_alf_param(alf_param_temp, cur_aps); } else { continue; } alf_reconstruct_coeff(state, alf_param_temp, CHANNEL_TYPE_CHROMA, true, true, arr_vars); //m_CABACEstimator->getCtx() = AlfCtx(ctxStart); memcpy(cabac_estimator, &ctx_start, sizeof(*cabac_estimator)); for (int comp_id = 1; comp_id < MAX_NUM_COMPONENT; comp_id++) { alf_cov_chroma = comp_id == COMPONENT_Cb ? alf_info->alf_covariance_u : alf_info->alf_covariance_v; alf_param_temp->enabled_flag[comp_id] = true; for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { double dist_unfilter_ctu = ctb_distortions_unfilter[comp_id][ctu_idx]; //cost on ctu_enable_flag[comp_id][ctu_idx] = 1; memcpy(&ctx_temp_start, cabac_estimator, sizeof(ctx_temp_start)); ctx_temp_start.only_count = 1; //rate //memcpy(&cabac_estimator, &ctx_temp_start, sizeof(cabac_estimator)); alf_cabac_reset_bits(cabac_estimator); //ctb flag code_alf_ctu_enable_flag(state, cabac_estimator, ctu_idx, comp_id, alf_param_temp); double rate_on = (23 - cabac_estimator->bits_left) + (cabac_estimator->num_buffered_bytes << 3); //frac_bits_scale*(double)838/*m_CABACEstimator->getEstFracBits()*/; const double ctu_lambda = lambda_chroma_weight > 0.0 ? 0/*cs.picture->m_uEnerHpCtu[ctbIdx]*/ / lambda_chroma_weight : lambda; double dist = MAX_DOUBLE; int num_alts = alf_param_temp->num_alternatives_chroma; //ctxTempBest = AlfCtx(m_CABACEstimator->getCtx()); memcpy(&ctx_temp_best, cabac_estimator, sizeof(ctx_temp_best)); ctx_temp_best.only_count = 1; double best_alt_rate = 0; double best_alt_cost = MAX_DOUBLE; int best_alt_idx = -1; //ctxTempAltStart = AlfCtx(ctxTempBest); memcpy(&ctx_temp_alt_start, &ctx_temp_best, sizeof(ctx_temp_alt_start)); for (int alt_idx = 0; alt_idx < num_alts; ++alt_idx) { if (alt_idx) { //m_CABACEstimator->getCtx() = AlfCtx(ctxTempAltStart); memcpy(cabac_estimator, &ctx_temp_alt_start, sizeof(*cabac_estimator)); } //m_CABACEstimator->resetBits(); alf_cabac_reset_bits(cabac_estimator); ctu_alternatives[comp_id][ctu_idx] = alt_idx; //m_CABACEstimator->codeAlfCtuAlternative(cs, ctbIdx, compId, &m_alfParamTemp); code_alf_ctu_alternative_ctu(state, cabac_estimator, ctu_idx, comp_id, alf_param_temp); double alt_rate = (23 - cabac_estimator->bits_left) + (cabac_estimator->num_buffered_bytes << 3); //frac_bits_scale * 0/*m_CABACEstimator->getEstFracBits()*/; double r_alt_cost = ctu_lambda * alt_rate; //distortion for (int i = 0; i < MAX_NUM_ALF_CHROMA_COEFF; i++) { filter_tmp[i] = arr_vars->chroma_coeff_final[alt_idx][i]; g_clip_tmp[i] = arr_vars->chroma_clipp_final[alt_idx][i]; } double alt_dist = calc_error_for_coeffs(&alf_cov_chroma[ctu_idx], g_clip_tmp, filter_tmp, MAX_NUM_ALF_CHROMA_COEFF, bit_depth); double alt_cost = alt_dist + r_alt_cost; if (alt_cost < best_alt_cost) { best_alt_cost = alt_cost; best_alt_idx = alt_idx; best_alt_rate = alt_rate; //ctxTempBest = AlfCtx(m_CABACEstimator->getCtx()); memcpy(&ctx_temp_best, cabac_estimator, sizeof(ctx_temp_best)); ctx_temp_best.only_count = 1; dist = alt_dist; } } ctu_alternatives[comp_id][ctu_idx] = best_alt_idx; rate_on += best_alt_rate; dist += dist_unfilter_ctu; //cost double cost_on = dist + ctu_lambda * rate_on; //cost off ctu_enable_flag[comp_id][ctu_idx] = 0; //rate memcpy(cabac_estimator, &ctx_temp_start, sizeof(*cabac_estimator)); alf_cabac_reset_bits(cabac_estimator); code_alf_ctu_enable_flag(state, cabac_estimator, ctu_idx, comp_id, alf_param_temp); //cost double cost_off = dist_unfilter_ctu + lambda * (23 - cabac_estimator->bits_left) + (cabac_estimator->num_buffered_bytes << 3); //frac_bits_scale*(double)838/*m_CABACEstimator->getEstFracBits()*/; if (cost_on < cost_off) { //m_CABACEstimator->getCtx() = AlfCtx(ctxTempBest); memcpy(cabac_estimator, &ctx_temp_best, sizeof(*cabac_estimator)); ctu_enable_flag[comp_id][ctu_idx] = 1; cur_cost += cost_on; } else { ctu_enable_flag[comp_id][ctu_idx] = 0; cur_cost += cost_off; } }//ctb_idx }//comp_id //chroma idc //setEnableFlag(m_alfSliceParamTemp, CHANNEL_TYPE_CHROMA, m_ctuEnableFlag); for (int comp_id = COMPONENT_Cb; comp_id <= COMPONENT_Cr; comp_id++) { alf_param_temp->enabled_flag[comp_id] = false; for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { if (ctu_enable_flag[comp_id][ctu_idx]) { alf_param_temp->enabled_flag[comp_id] = true; break; } } } if (cur_cost < cost_min) { cost_min = cur_cost; state->slice->alf->tile_group_chroma_aps_id = cur_aps_id; state->slice->alf->tile_group_alf_enabled_flag[COMPONENT_Cb] = alf_param_temp->enabled_flag[COMPONENT_Cb]; state->slice->alf->tile_group_alf_enabled_flag[COMPONENT_Cr] = alf_param_temp->enabled_flag[COMPONENT_Cr]; copy_ctu_enable_flag(ctu_enable_flag_tmp, ctu_enable_flag, CHANNEL_TYPE_CHROMA, num_ctus_in_pic); for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { ctu_alternatives_tmp[COMPONENT_Cb][ctu_idx] = ctu_alternatives[COMPONENT_Cb][ctu_idx]; ctu_alternatives_tmp[COMPONENT_Cr][ctu_idx] = ctu_alternatives[COMPONENT_Cr][ctu_idx]; } } } if (new_aps_id_chroma >= 0) { state->slice->alf->tile_group_cc_alf_cb_aps_id = new_aps_id_chroma; state->slice->alf->tile_group_cc_alf_cr_aps_id = new_aps_id_chroma; } if (cost_off < cost_min) { state->slice->alf->tile_group_alf_enabled_flag[COMPONENT_Cb] = false; state->slice->alf->tile_group_alf_enabled_flag[COMPONENT_Cr] = false; set_ctu_enable_flag(ctu_enable_flag, CHANNEL_TYPE_CHROMA, 0, num_ctus_in_pic); } else { copy_ctu_enable_flag(ctu_enable_flag, ctu_enable_flag_tmp, CHANNEL_TYPE_CHROMA, num_ctus_in_pic); for (int ctu_idx = 0; ctu_idx < num_ctus_in_pic; ctu_idx++) { ctu_alternatives[COMPONENT_Cb][ctu_idx] = ctu_alternatives_tmp[COMPONENT_Cb][ctu_idx]; ctu_alternatives[COMPONENT_Cr][ctu_idx] = ctu_alternatives_tmp[COMPONENT_Cr][ctu_idx]; } if (state->slice->alf->tile_group_chroma_aps_id == new_aps_id_chroma) //new filter { //APS* newAPS = m_apsMap->getPS(new_aps_id_chroma); alf_aps* new_aps = &state->tile->frame->alf_param_set_map[new_aps_id_chroma + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set; if (new_aps->aps_id < 0 || new_aps->aps_id >= ALF_CTB_MAX_NUM_APS) //new_aps == NULL { //newAPS = m_apsMap->allocatePS(new_aps_id); assert(new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS < ALF_CTB_MAX_NUM_APS); //Invalid PS id bool found = false; for (int i = 0; i < (sizeof(state->tile->frame->alf_param_set_map) / sizeof(state->tile->frame->alf_param_set_map[0])); i++) { if (state->tile->frame->alf_param_set_map[i].parameter_set.aps_id == new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS) { found = true; } } if (!found) { state->tile->frame->alf_param_set_map[new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].b_changed = true; //state->tile->frame->alf_param_set_map[new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].p_nalu_data = 0; //state->tile->frame->alf_param_set_map[new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set = malloc(sizeof(alf_aps)); state->tile->frame->alf_param_set_map[new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set.aps_id = new_aps_id + T_ALF_APS; } //copy_alf_param(new_aps, &state->tile->frame->alf_param_set_map[new_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set); new_aps->aps_id = new_aps_id; new_aps->aps_type = T_ALF_APS; reset_alf_param(new_aps); } new_aps->new_filter_flag[CHANNEL_TYPE_CHROMA] = true; if (!alf_aps_new_filters_best.new_filter_flag[CHANNEL_TYPE_LUMA]) { new_aps->new_filter_flag[CHANNEL_TYPE_LUMA] = false; } new_aps->num_alternatives_chroma = aps->num_alternatives_chroma; new_aps->non_linear_flag[CHANNEL_TYPE_CHROMA] = aps->non_linear_flag[CHANNEL_TYPE_CHROMA]; new_aps->temporal_id = state->slice->id; for (int alt_idx = 0; alt_idx < MAX_NUM_ALF_ALTERNATIVES_CHROMA; ++alt_idx) { for (int i = 0; i < MAX_NUM_ALF_CHROMA_COEFF; i++) { new_aps->chroma_coeff[alt_idx][i] = aps->chroma_coeff[alt_idx][i]; new_aps->chroma_clipp[alt_idx][i] = aps->chroma_clipp[alt_idx][i]; } } state->tile->frame->alf_param_set_map[new_aps_id_chroma + NUM_APS_TYPE_LEN + T_ALF_APS].b_changed = true; alf_info->aps_id_start = new_aps_id_chroma; } apss[state->slice->alf->tile_group_chroma_aps_id].aps_id = state->tile->frame->alf_param_set_map[state->slice->alf->tile_group_chroma_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set.aps_id; apss[state->slice->alf->tile_group_chroma_aps_id].aps_type = state->tile->frame->alf_param_set_map[state->slice->alf->tile_group_chroma_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set.aps_type; copy_alf_param(&apss[state->slice->alf->tile_group_chroma_aps_id], &state->tile->frame->alf_param_set_map[state->slice->alf->tile_group_chroma_aps_id + NUM_APS_TYPE_LEN + T_ALF_APS].parameter_set); } } } static void alf_reconstruct(encoder_state_t * const state, array_variables *arr_vars) { if (!state->slice->alf->tile_group_alf_enabled_flag[COMPONENT_Y]) { return; } alf_reconstruct_coeff_aps(state, true, state->slice->alf->tile_group_alf_enabled_flag[COMPONENT_Cb] || state->slice->alf->tile_group_alf_enabled_flag[COMPONENT_Cr], false, arr_vars); alf_info_t *alf_info = state->tile->frame->alf_info; bool **ctu_enable_flags = alf_info->ctu_enable_flag; enum kvz_chroma_format chroma_fmt = state->encoder_control->chroma_format; bool chroma_scale_x = (chroma_fmt == KVZ_CSP_444) ? 0 : 1; bool chroma_scale_y = (chroma_fmt != KVZ_CSP_420) ? 0 : 1; const int alf_vb_luma_ctu_height = LCU_WIDTH; const int alf_vb_chma_ctu_height = (LCU_WIDTH >> ((chroma_fmt == KVZ_CSP_420) ? 1 : 0)); const int alf_vb_luma_pos = LCU_WIDTH - ALF_VB_POS_ABOVE_CTUROW_LUMA; const int alf_vb_chma_pos = (LCU_WIDTH >> ((chroma_fmt == KVZ_CSP_420) ? 1 : 0)) - ALF_VB_POS_ABOVE_CTUROW_CHMA; const int luma_height = state->tile->frame->height; const int luma_width = state->tile->frame->width; const int max_cu_width = LCU_WIDTH; const int max_cu_height = LCU_WIDTH; int ctu_idx = 0; const int luma_stride = state->tile->frame->rec->stride; const int chroma_stride = luma_stride >> chroma_scale_x; const int chroma_height = luma_height >> chroma_scale_y; const int chroma_padding = MAX_ALF_PADDING_SIZE >> chroma_scale_x; const int index_luma = -(luma_stride * MAX_ALF_PADDING_SIZE + MAX_ALF_PADDING_SIZE); const int index_chroma = -(chroma_stride * chroma_padding + chroma_padding); //Copy reconstructed samples to a buffer. memcpy(&alf_info->alf_tmp_y[index_luma], &state->tile->frame->rec->y[index_luma], sizeof(kvz_pixel) * luma_stride * (luma_height + MAX_ALF_PADDING_SIZE * 2)); memcpy(&alf_info->alf_tmp_u[index_chroma], &state->tile->frame->rec->u[index_chroma], sizeof(kvz_pixel) * chroma_stride * (chroma_height + chroma_padding * 2)); memcpy(&alf_info->alf_tmp_v[index_chroma], &state->tile->frame->rec->v[index_chroma], sizeof(kvz_pixel) * chroma_stride * (chroma_height + chroma_padding * 2)); for (int y_pos = 0; y_pos < luma_height; y_pos += max_cu_height) { for (int x_pos = 0; x_pos < luma_width; x_pos += max_cu_width) { const int width = (x_pos + max_cu_width > luma_width) ? (luma_width - x_pos) : max_cu_width; const int height = (y_pos + max_cu_height > luma_height) ? (luma_height - y_pos) : max_cu_height; bool ctu_enable_flag = ctu_enable_flags[COMPONENT_Y][ctu_idx]; for (int comp_idx = 1; comp_idx < MAX_NUM_COMPONENT; comp_idx++) { ctu_enable_flag |= ctu_enable_flags[comp_idx][ctu_idx] > 0; } { if (ctu_enable_flags[COMPONENT_Y][ctu_idx]) { short filter_set_index = alf_info->alf_ctb_filter_index[ctu_idx]; short *coeff; int16_t *clip; if (filter_set_index >= ALF_NUM_FIXED_FILTER_SETS) { coeff = arr_vars->coeff_aps_luma[filter_set_index - ALF_NUM_FIXED_FILTER_SETS]; clip = arr_vars->clipp_aps_luma[filter_set_index - ALF_NUM_FIXED_FILTER_SETS]; } else { coeff = arr_vars->fixed_filter_set_coeff_dec[filter_set_index]; clip = arr_vars->clip_default; } kvz_alf_filter_7x7_blk(state, alf_info->alf_tmp_y, state->tile->frame->rec->y, luma_stride, luma_stride, coeff, clip, arr_vars->clp_rngs.comp[COMPONENT_Y], width, height, x_pos, y_pos, x_pos, y_pos, alf_vb_luma_pos, alf_vb_luma_ctu_height); } for (int comp_idx = 1; comp_idx < MAX_NUM_COMPONENT; comp_idx++) { alf_component_id comp_id = comp_idx; if (ctu_enable_flags[comp_idx][ctu_idx]) { kvz_pixel *dst_pixels = comp_id - 1 ? state->tile->frame->rec->v : state->tile->frame->rec->u; const kvz_pixel *src_pixels = comp_id - 1 ? alf_info->alf_tmp_v : alf_info->alf_tmp_u; const int alt_num = alf_info->ctu_alternative[comp_id][ctu_idx]; kvz_alf_filter_5x5_blk(state, src_pixels, dst_pixels, chroma_stride, chroma_stride, arr_vars->chroma_coeff_final[alt_num], arr_vars->chroma_clipp_final[alt_num], arr_vars->clp_rngs.comp[comp_idx], width >> chroma_scale_x, height >> chroma_scale_y, x_pos >> chroma_scale_x, y_pos >> chroma_scale_y, x_pos >> chroma_scale_x, y_pos >> chroma_scale_y, alf_vb_chma_pos, alf_vb_chma_ctu_height); } } } ctu_idx++; } } } static void alf_derive_classification(encoder_state_t * const state, const int width, const int height, int x_pos, int y_pos, const int blk_dst_x, const int blk_dst_y) { enum kvz_chroma_format chroma_fmt = state->encoder_control->chroma_format; bool chroma_scale_x = (chroma_fmt == KVZ_CSP_444) ? 0 : 1; bool chroma_scale_y = (chroma_fmt != KVZ_CSP_420) ? 0 : 1; const int alf_vb_luma_ctu_height = LCU_WIDTH; const int alf_vb_luma_pos = LCU_WIDTH - ALF_VB_POS_ABOVE_CTUROW_LUMA; int32_t pic_height = state->tile->frame->rec->height; int32_t pic_width = state->tile->frame->rec->width; int max_height = y_pos + height; int max_width = x_pos + width; adjust_pixels(state->tile->frame->rec->y, x_pos, pic_width, y_pos, pic_height, state->tile->frame->rec->stride, pic_width, pic_height); adjust_pixels_chroma(state->tile->frame->rec->u, x_pos >> chroma_scale_x, pic_width >> chroma_scale_x, y_pos >> chroma_scale_y, pic_height >> chroma_scale_y, state->tile->frame->rec->stride >> chroma_scale_x, pic_width >> chroma_scale_x, pic_height >> chroma_scale_y); adjust_pixels_chroma(state->tile->frame->rec->v, x_pos >> chroma_scale_x, pic_width >> chroma_scale_x, y_pos >> chroma_scale_y, pic_height >> chroma_scale_y, state->tile->frame->rec->stride >> chroma_scale_x, pic_width >> chroma_scale_x, pic_height >> chroma_scale_y); for (int i = y_pos; i < max_height; i += CLASSIFICATION_BLK_SIZE) { int n_height = MIN(i + CLASSIFICATION_BLK_SIZE, max_height) - i; for (int j = x_pos; j < max_width; j += CLASSIFICATION_BLK_SIZE) { int n_width = MIN(j + CLASSIFICATION_BLK_SIZE, max_width) - j; kvz_alf_derive_classification_blk(state, state->encoder_control->cfg.input_bitdepth + 4, n_height, n_width, j, i, j - x_pos + blk_dst_x, i - y_pos + blk_dst_y, alf_vb_luma_ctu_height, alf_vb_luma_pos); } } } void kvz_alf_enc_process(encoder_state_t *const state) { alf_init_covariance(state->tile->frame, state->encoder_control->chroma_format); alf_info_t *alf_info = state->tile->frame->alf_info; alf_create_frame_buffer(state, alf_info); /* //if (!layerIdx && cs.slice->getPendingRasInit() if (1 && (false || (state->frame->pictype == KVZ_NAL_IDR_W_RADL || state->frame->pictype == KVZ_NAL_IDR_N_LP))) { for (int i = 0; i < ALF_CTB_MAX_NUM_APS; i++) { reset_aps(&state->slice->alf->apss[i], state->encoder_control->cfg.alf_type == KVZ_ALF_FULL); if (state->tile->frame->alf_param_set_map[i + T_ALF_APS].b_changed) { alf_aps* alf_aps = &state->tile->frame->alf_param_set_map[i + T_ALF_APS].parameter_set; state->tile->frame->alf_param_set_map[i + T_ALF_APS].b_changed = false; reset_aps(alf_aps, state->encoder_control->cfg.alf_type == KVZ_ALF_FULL); } } alf_info->aps_id_start = ALF_CTB_MAX_NUM_APS; }*/ alf_aps alf_param; reset_alf_param(&alf_param); cc_alf_filter_param *cc_filter_param = state->slice->alf->cc_filter_param; enum kvz_chroma_format chroma_fmt = state->encoder_control->chroma_format; bool chroma_scale_x = (chroma_fmt == KVZ_CSP_444) ? 0 : 1; bool chroma_scale_y = (chroma_fmt != KVZ_CSP_420) ? 0 : 1; int8_t kvz_bit_depth = state->encoder_control->bitdepth; const int32_t num_ctus_in_pic = state->tile->frame->width_in_lcu * state->tile->frame->height_in_lcu; const int8_t input_bitdepth = state->encoder_control->bitdepth; double lambda_chroma_weight = 0.0; cabac_data_t ctx_start; cabac_data_t ctx_start_cc_alf; cabac_data_t *cabac_estimator = &alf_info->cabac_estimator; memcpy(cabac_estimator, &state->cabac, sizeof(*cabac_estimator)); memcpy(&ctx_start, &state->cabac, sizeof(ctx_start)); memcpy(&ctx_start_cc_alf, cabac_estimator, sizeof(ctx_start_cc_alf)); cabac_estimator->only_count = 1; ctx_start.only_count = 1; ctx_start_cc_alf.only_count = 1; // derive classification const int luma_height = state->tile->frame->height; const int luma_width = state->tile->frame->width; array_variables arr_vars; bool init_values = false; if (!init_values) { assert(MAX_ALF_NUM_CLIPPING_VALUES > 0); //"g_alf_num_clipping_values[CHANNEL_TYPE_LUMA] must be at least one" arr_vars.alf_clipping_values[CHANNEL_TYPE_LUMA][0] = 1 << input_bitdepth; int shift_luma = input_bitdepth - 8; for (int i = 1; i < MAX_ALF_NUM_CLIPPING_VALUES; ++i) { arr_vars.alf_clipping_values[CHANNEL_TYPE_LUMA][i] = 1 << (7 - 2 * i + shift_luma); } assert(MAX_ALF_NUM_CLIPPING_VALUES > 0); //"g_alf_num_clipping_values[CHANNEL_TYPE_CHROMA] must be at least one" arr_vars.alf_clipping_values[CHANNEL_TYPE_CHROMA][0] = 1 << input_bitdepth; int shift_chroma = input_bitdepth - 8; for (int i = 1; i < MAX_ALF_NUM_CLIPPING_VALUES; ++i) { arr_vars.alf_clipping_values[CHANNEL_TYPE_CHROMA][i] = 1 << (7 - 2 * i + shift_chroma); } for (int i = 0; i < MAX_NUM_ALF_LUMA_COEFF * MAX_NUM_ALF_CLASSES; i++) { arr_vars.clip_default[i] = arr_vars.alf_clipping_values[CHANNEL_TYPE_LUMA][0]; } for (int filter_set_index = 0; filter_set_index < ALF_NUM_FIXED_FILTER_SETS; filter_set_index++) { for (int class_idx = 0; class_idx < MAX_NUM_ALF_CLASSES; class_idx++) { int fixed_filter_idx = g_class_to_filter_mapping[filter_set_index][class_idx]; for (int i = 0; i < MAX_NUM_ALF_LUMA_COEFF - 1; i++) { arr_vars.fixed_filter_set_coeff_dec[filter_set_index][class_idx * MAX_NUM_ALF_LUMA_COEFF + i] = g_fixed_filter_set_coeff[fixed_filter_idx][i]; } arr_vars.fixed_filter_set_coeff_dec[filter_set_index][class_idx * MAX_NUM_ALF_LUMA_COEFF + MAX_NUM_ALF_LUMA_COEFF - 1] = (1 << (input_bitdepth - 1)); } } //Default clp_rng arr_vars.clp_rngs.comp[COMPONENT_Y].min = arr_vars.clp_rngs.comp[COMPONENT_Cb].min = arr_vars.clp_rngs.comp[COMPONENT_Cr].min = 0; arr_vars.clp_rngs.comp[COMPONENT_Y].max = (1 << kvz_bit_depth) - 1; arr_vars.clp_rngs.comp[COMPONENT_Y].bd = kvz_bit_depth; arr_vars.clp_rngs.comp[COMPONENT_Y].n = 0; arr_vars.clp_rngs.comp[COMPONENT_Cb].max = arr_vars.clp_rngs.comp[COMPONENT_Cr].max = (1 << kvz_bit_depth) - 1; arr_vars.clp_rngs.comp[COMPONENT_Cb].bd = arr_vars.clp_rngs.comp[COMPONENT_Cr].bd = kvz_bit_depth; arr_vars.clp_rngs.comp[COMPONENT_Cb].n = arr_vars.clp_rngs.comp[COMPONENT_Cr].n = 0; arr_vars.clp_rngs.used = arr_vars.clp_rngs.chroma = false; init_values = true; } for (int y_pos = 0; y_pos < luma_height; y_pos += LCU_WIDTH) { for (int x_pos = 0; x_pos < luma_width; x_pos += LCU_WIDTH) { const int width = (x_pos + LCU_WIDTH > luma_width) ? (luma_width - x_pos) : LCU_WIDTH; const int height = (y_pos + LCU_WIDTH > luma_height) ? (luma_height - y_pos) : LCU_WIDTH; { alf_derive_classification(state, width, height, x_pos, y_pos, x_pos, y_pos); } } } // get CTB stats for filtering alf_derive_stats_for_filtering(state, arr_vars.alf_clipping_values); for (int ctb_iIdx = 0; ctb_iIdx < num_ctus_in_pic; ctb_iIdx++) { alf_info->alf_ctb_filter_index[ctb_iIdx] = ALF_NUM_FIXED_FILTER_SETS; } // consider using new filter (only) alf_param.new_filter_flag[CHANNEL_TYPE_LUMA] = true; alf_param.new_filter_flag[CHANNEL_TYPE_CHROMA] = true; state->slice->alf->tile_group_num_aps = 1; // Only new filter for RD cost optimization // derive filter (luma) alf_encoder(state, &alf_param, CHANNEL_TYPE_LUMA, lambda_chroma_weight, &arr_vars ); // derive filter (chroma) if (state->encoder_control->chroma_format != KVZ_CSP_400) { alf_encoder(state, &alf_param, CHANNEL_TYPE_CHROMA, lambda_chroma_weight, &arr_vars ); } // let alfEncoderCtb decide now alf_param.new_filter_flag[CHANNEL_TYPE_LUMA] = false; alf_param.new_filter_flag[CHANNEL_TYPE_CHROMA] = false; state->slice->alf->tile_group_num_aps = 0; //m_CABACEstimator->getCtx() = AlfCtx(ctxStart); memcpy(cabac_estimator, &ctx_start, sizeof(*cabac_estimator)); alf_encoder_ctb(state, &alf_param, lambda_chroma_weight, &arr_vars); //for (int s = 0; s < state.; s++) //numSliceSegments { if (state->encoder_control->cfg.lossless) { for (uint32_t ctb_idx = 0; ctb_idx < num_ctus_in_pic; ctb_idx++) //pcPic->slices[s]->getNumCtuInSlice() { //uint32_t ctuRsAddr = pcPic->slices[s]->getCtuAddrInSlice(ctuIdx); state->tile->frame->alf_info->ctu_enable_flag[COMPONENT_Y][ctb_idx] = 0; state->tile->frame->alf_info->ctu_enable_flag[COMPONENT_Cb][ctb_idx] = 0; state->tile->frame->alf_info->ctu_enable_flag[COMPONENT_Cr][ctb_idx] = 0; } } } alf_reconstruct(state, &arr_vars); if (state->encoder_control->cfg.alf_type != KVZ_ALF_FULL) { return; } // Do not transmit CC ALF if it is unchanged if (state->slice->alf->tile_group_alf_enabled_flag[COMPONENT_Y]) { for (int32_t luma_alf_aps_id = 0; luma_alf_aps_id < state->slice->alf->tile_group_num_aps; luma_alf_aps_id++) { //APS* aps = (luma_alf_aps_id >= 0) ? m_apsMap->getPS((luma_alf_aps_id << NUM_APS_TYPE_LEN) + ALF_APS) : nullptr; int aps_id = state->slice->alf->tile_group_luma_aps_id[luma_alf_aps_id]; alf_aps* aps = (aps_id >= 0) ? &state->tile->frame->alf_param_set_map[aps_id + T_ALF_APS].parameter_set : NULL; bool changed = state->tile->frame->alf_param_set_map[aps_id + T_ALF_APS].b_changed; if (aps && changed) { aps->cc_alf_aps_param.new_cc_alf_filter[0] = false; aps->cc_alf_aps_param.new_cc_alf_filter[1] = false; } } } int chroma_alf_aps_id = (state->slice->alf->tile_group_alf_enabled_flag[COMPONENT_Cb] || state->slice->alf->tile_group_alf_enabled_flag[COMPONENT_Cr]) ? state->slice->alf->tile_group_chroma_aps_id : -1; alf_aps* aps = (chroma_alf_aps_id >= 0) ? &state->tile->frame->alf_param_set_map[chroma_alf_aps_id + T_ALF_APS].parameter_set : NULL; bool changed = (chroma_alf_aps_id >= 0) ? state->tile->frame->alf_param_set_map[chroma_alf_aps_id + T_ALF_APS].b_changed : 0; if (aps && changed) { aps->cc_alf_aps_param.new_cc_alf_filter[0] = false; aps->cc_alf_aps_param.new_cc_alf_filter[1] = false; } const kvz_picture *org_yuv = state->tile->frame->source; const kvz_picture *rec_yuv = state->tile->frame->rec; const int luma_stride = state->tile->frame->rec->stride; const int chroma_stride = luma_stride >> chroma_scale_x; const int chroma_height = luma_height >> chroma_scale_y; const int chroma_padding = MAX_ALF_PADDING_SIZE >> chroma_scale_x; const int index_chroma = -(chroma_stride * chroma_padding + chroma_padding); //Copy reconstructed samples to a buffer. memcpy(&alf_info->alf_tmp_u[index_chroma], &state->tile->frame->rec->u[index_chroma], sizeof(kvz_pixel) * chroma_stride * (chroma_height + chroma_padding * 2)); memcpy(&alf_info->alf_tmp_v[index_chroma], &state->tile->frame->rec->v[index_chroma], sizeof(kvz_pixel) * chroma_stride * (chroma_height + chroma_padding * 2)); adjust_pixels_chroma(alf_info->alf_tmp_u, 0, rec_yuv->width >> chroma_scale_x, 0, rec_yuv->height >> chroma_scale_y, rec_yuv->stride >> chroma_scale_x, rec_yuv->width >> chroma_scale_x, rec_yuv->height >> chroma_scale_y); adjust_pixels_chroma(alf_info->alf_tmp_v, 0, rec_yuv->width >> chroma_scale_x, 0, rec_yuv->height >> chroma_scale_y, rec_yuv->stride >> chroma_scale_x, rec_yuv->width >> chroma_scale_x, rec_yuv->height >> chroma_scale_y); const int num_ctus_in_width = state->tile->frame->width_in_lcu; derive_stats_for_cc_alf_filtering(state, org_yuv, COMPONENT_Cb, num_ctus_in_width, (0 + 1)); derive_stats_for_cc_alf_filtering(state, org_yuv, COMPONENT_Cr, num_ctus_in_width, (0 + 1)); init_distortion_cc_alf(alf_info->alf_covariance_cc_alf, alf_info->ctb_distortion_unfilter, num_ctus_in_pic); memcpy(cabac_estimator, &ctx_start_cc_alf, sizeof(*cabac_estimator)); derive_cc_alf_filter(state, COMPONENT_Cb, org_yuv, rec_yuv, arr_vars.cc_reuse_aps_id); memcpy(cabac_estimator, &ctx_start_cc_alf, sizeof(*cabac_estimator)); derive_cc_alf_filter(state, COMPONENT_Cr, org_yuv, rec_yuv, arr_vars.cc_reuse_aps_id); setup_cc_alf_aps(state, arr_vars.cc_reuse_aps_id); for (alf_component_id comp_idx = 1; comp_idx < (state->encoder_control->chroma_format == KVZ_CSP_400 ? 1 : MAX_NUM_COMPONENT); comp_idx++) { if (cc_filter_param->cc_alf_filter_enabled[comp_idx - 1]) { kvz_pixel* rec_uv = comp_idx == COMPONENT_Cb ? rec_yuv->u : rec_yuv->v; const int luma_stride = rec_yuv->stride; apply_cc_alf_filter(state, comp_idx, rec_uv, alf_info->alf_tmp_y, luma_stride, alf_info->cc_alf_filter_control[comp_idx - 1], cc_filter_param->cc_alf_coeff[comp_idx - 1], -1, &arr_vars); } } alf_covariance_destroy(state->tile->frame); }