1 /**************************************************************************
2 *
3 * Copyright 2017 Advanced Micro Devices, Inc.
4 *
5 * SPDX-License-Identifier: MIT
6 *
7 **************************************************************************/
8
9 #include <stdint.h>
10 #include "util/u_math.h"
11
12 #include "ac_vcn_dec.h"
13 #include "ac_vcn_av1_default.h"
14
15 static unsigned
ac_vcn_dec_frame_ctx_size_av1(unsigned av1_version)16 ac_vcn_dec_frame_ctx_size_av1(unsigned av1_version)
17 {
18 return av1_version == RDECODE_AV1_VER_0
19 ? align(sizeof(rvcn_av1_frame_context_t), 2048)
20 : align(sizeof(rvcn_av1_vcn4_frame_context_t), 2048);
21 }
22
23 unsigned
ac_vcn_dec_calc_ctx_size_av1(unsigned av1_version)24 ac_vcn_dec_calc_ctx_size_av1(unsigned av1_version)
25 {
26 unsigned frame_ctxt_size = ac_vcn_dec_frame_ctx_size_av1(av1_version);
27 unsigned ctx_size = (9 + 4) * frame_ctxt_size + 9 * 64 * 34 * 512 + 9 * 64 * 34 * 256 * 5;
28
29 int num_64x64_CTB_8k = 68;
30 int num_128x128_CTB_8k = 34;
31 int sdb_pitch_64x64 = align(32 * num_64x64_CTB_8k, 256) * 2;
32 int sdb_pitch_128x128 = align(32 * num_128x128_CTB_8k, 256) * 2;
33 int sdb_lf_size_ctb_64x64 = sdb_pitch_64x64 * (align(1728, 64) / 64);
34 int sdb_lf_size_ctb_128x128 = sdb_pitch_128x128 * (align(3008, 64) / 64);
35 int sdb_superres_size_ctb_64x64 = sdb_pitch_64x64 * (align(3232, 64) / 64);
36 int sdb_superres_size_ctb_128x128 = sdb_pitch_128x128 * (align(6208, 64) / 64);
37 int sdb_output_size_ctb_64x64 = sdb_pitch_64x64 * (align(1312, 64) / 64);
38 int sdb_output_size_ctb_128x128 = sdb_pitch_128x128 * (align(2336, 64) / 64);
39 int sdb_fg_avg_luma_size_ctb_64x64 = sdb_pitch_64x64 * (align(384, 64) / 64);
40 int sdb_fg_avg_luma_size_ctb_128x128 = sdb_pitch_128x128 * (align(640, 64) / 64);
41
42 ctx_size += (MAX2(sdb_lf_size_ctb_64x64, sdb_lf_size_ctb_128x128) +
43 MAX2(sdb_superres_size_ctb_64x64, sdb_superres_size_ctb_128x128) +
44 MAX2(sdb_output_size_ctb_64x64, sdb_output_size_ctb_128x128) +
45 MAX2(sdb_fg_avg_luma_size_ctb_64x64, sdb_fg_avg_luma_size_ctb_128x128)) *
46 2 +
47 68 * 512;
48
49 return ctx_size;
50 }
51
52 static void
ac_vcn_av1_init_mode_probs(void * prob)53 ac_vcn_av1_init_mode_probs(void *prob)
54 {
55 rvcn_av1_frame_context_t *fc = (rvcn_av1_frame_context_t *)prob;
56 int i;
57
58 memcpy(fc->palette_y_size_cdf, default_palette_y_size_cdf, sizeof(default_palette_y_size_cdf));
59 memcpy(fc->palette_uv_size_cdf, default_palette_uv_size_cdf, sizeof(default_palette_uv_size_cdf));
60 memcpy(fc->palette_y_color_index_cdf, default_palette_y_color_index_cdf, sizeof(default_palette_y_color_index_cdf));
61 memcpy(fc->palette_uv_color_index_cdf, default_palette_uv_color_index_cdf,
62 sizeof(default_palette_uv_color_index_cdf));
63 memcpy(fc->kf_y_cdf, default_kf_y_mode_cdf, sizeof(default_kf_y_mode_cdf));
64 memcpy(fc->angle_delta_cdf, default_angle_delta_cdf, sizeof(default_angle_delta_cdf));
65 memcpy(fc->comp_inter_cdf, default_comp_inter_cdf, sizeof(default_comp_inter_cdf));
66 memcpy(fc->comp_ref_type_cdf, default_comp_ref_type_cdf, sizeof(default_comp_ref_type_cdf));
67 memcpy(fc->uni_comp_ref_cdf, default_uni_comp_ref_cdf, sizeof(default_uni_comp_ref_cdf));
68 memcpy(fc->palette_y_mode_cdf, default_palette_y_mode_cdf, sizeof(default_palette_y_mode_cdf));
69 memcpy(fc->palette_uv_mode_cdf, default_palette_uv_mode_cdf, sizeof(default_palette_uv_mode_cdf));
70 memcpy(fc->comp_ref_cdf, default_comp_ref_cdf, sizeof(default_comp_ref_cdf));
71 memcpy(fc->comp_bwdref_cdf, default_comp_bwdref_cdf, sizeof(default_comp_bwdref_cdf));
72 memcpy(fc->single_ref_cdf, default_single_ref_cdf, sizeof(default_single_ref_cdf));
73 memcpy(fc->txfm_partition_cdf, default_txfm_partition_cdf, sizeof(default_txfm_partition_cdf));
74 memcpy(fc->compound_index_cdf, default_compound_idx_cdfs, sizeof(default_compound_idx_cdfs));
75 memcpy(fc->comp_group_idx_cdf, default_comp_group_idx_cdfs, sizeof(default_comp_group_idx_cdfs));
76 memcpy(fc->newmv_cdf, default_newmv_cdf, sizeof(default_newmv_cdf));
77 memcpy(fc->zeromv_cdf, default_zeromv_cdf, sizeof(default_zeromv_cdf));
78 memcpy(fc->refmv_cdf, default_refmv_cdf, sizeof(default_refmv_cdf));
79 memcpy(fc->drl_cdf, default_drl_cdf, sizeof(default_drl_cdf));
80 memcpy(fc->motion_mode_cdf, default_motion_mode_cdf, sizeof(default_motion_mode_cdf));
81 memcpy(fc->obmc_cdf, default_obmc_cdf, sizeof(default_obmc_cdf));
82 memcpy(fc->inter_compound_mode_cdf, default_inter_compound_mode_cdf, sizeof(default_inter_compound_mode_cdf));
83 memcpy(fc->compound_type_cdf, default_compound_type_cdf, sizeof(default_compound_type_cdf));
84 memcpy(fc->wedge_idx_cdf, default_wedge_idx_cdf, sizeof(default_wedge_idx_cdf));
85 memcpy(fc->interintra_cdf, default_interintra_cdf, sizeof(default_interintra_cdf));
86 memcpy(fc->wedge_interintra_cdf, default_wedge_interintra_cdf, sizeof(default_wedge_interintra_cdf));
87 memcpy(fc->interintra_mode_cdf, default_interintra_mode_cdf, sizeof(default_interintra_mode_cdf));
88 memcpy(fc->pred_cdf, default_segment_pred_cdf, sizeof(default_segment_pred_cdf));
89 memcpy(fc->switchable_restore_cdf, default_switchable_restore_cdf, sizeof(default_switchable_restore_cdf));
90 memcpy(fc->wiener_restore_cdf, default_wiener_restore_cdf, sizeof(default_wiener_restore_cdf));
91 memcpy(fc->sgrproj_restore_cdf, default_sgrproj_restore_cdf, sizeof(default_sgrproj_restore_cdf));
92 memcpy(fc->y_mode_cdf, default_if_y_mode_cdf, sizeof(default_if_y_mode_cdf));
93 memcpy(fc->uv_mode_cdf, default_uv_mode_cdf, sizeof(default_uv_mode_cdf));
94 memcpy(fc->switchable_interp_cdf, default_switchable_interp_cdf, sizeof(default_switchable_interp_cdf));
95 memcpy(fc->partition_cdf, default_partition_cdf, sizeof(default_partition_cdf));
96 memcpy(fc->intra_ext_tx_cdf, default_intra_ext_tx_cdf, sizeof(default_intra_ext_tx_cdf));
97 memcpy(fc->inter_ext_tx_cdf, default_inter_ext_tx_cdf, sizeof(default_inter_ext_tx_cdf));
98 memcpy(fc->skip_cdfs, default_skip_cdfs, sizeof(default_skip_cdfs));
99 memcpy(fc->intra_inter_cdf, default_intra_inter_cdf, sizeof(default_intra_inter_cdf));
100 memcpy(fc->tree_cdf, default_seg_tree_cdf, sizeof(default_seg_tree_cdf));
101 for (i = 0; i < SPATIAL_PREDICTION_PROBS; ++i)
102 memcpy(fc->spatial_pred_seg_cdf[i], default_spatial_pred_seg_tree_cdf[i],
103 sizeof(default_spatial_pred_seg_tree_cdf[i]));
104 memcpy(fc->tx_size_cdf, default_tx_size_cdf, sizeof(default_tx_size_cdf));
105 memcpy(fc->delta_q_cdf, default_delta_q_cdf, sizeof(default_delta_q_cdf));
106 memcpy(fc->skip_mode_cdfs, default_skip_mode_cdfs, sizeof(default_skip_mode_cdfs));
107 memcpy(fc->delta_lf_cdf, default_delta_lf_cdf, sizeof(default_delta_lf_cdf));
108 memcpy(fc->delta_lf_multi_cdf, default_delta_lf_multi_cdf, sizeof(default_delta_lf_multi_cdf));
109 memcpy(fc->cfl_sign_cdf, default_cfl_sign_cdf, sizeof(default_cfl_sign_cdf));
110 memcpy(fc->cfl_alpha_cdf, default_cfl_alpha_cdf, sizeof(default_cfl_alpha_cdf));
111 memcpy(fc->filter_intra_cdfs, default_filter_intra_cdfs, sizeof(default_filter_intra_cdfs));
112 memcpy(fc->filter_intra_mode_cdf, default_filter_intra_mode_cdf, sizeof(default_filter_intra_mode_cdf));
113 memcpy(fc->intrabc_cdf, default_intrabc_cdf, sizeof(default_intrabc_cdf));
114 }
115
116 static void
ac_vcn_av1_init_mv_probs(void * prob)117 ac_vcn_av1_init_mv_probs(void *prob)
118 {
119 rvcn_av1_frame_context_t *fc = (rvcn_av1_frame_context_t *)prob;
120
121 memcpy(fc->nmvc_joints_cdf, default_nmv_context.joints_cdf, sizeof(default_nmv_context.joints_cdf));
122 memcpy(fc->nmvc_0_bits_cdf, default_nmv_context.comps[0].bits_cdf, sizeof(default_nmv_context.comps[0].bits_cdf));
123 memcpy(fc->nmvc_0_class0_cdf, default_nmv_context.comps[0].class0_cdf,
124 sizeof(default_nmv_context.comps[0].class0_cdf));
125 memcpy(fc->nmvc_0_class0_fp_cdf, default_nmv_context.comps[0].class0_fp_cdf,
126 sizeof(default_nmv_context.comps[0].class0_fp_cdf));
127 memcpy(fc->nmvc_0_class0_hp_cdf, default_nmv_context.comps[0].class0_hp_cdf,
128 sizeof(default_nmv_context.comps[0].class0_hp_cdf));
129 memcpy(fc->nmvc_0_classes_cdf, default_nmv_context.comps[0].classes_cdf,
130 sizeof(default_nmv_context.comps[0].classes_cdf));
131 memcpy(fc->nmvc_0_fp_cdf, default_nmv_context.comps[0].fp_cdf, sizeof(default_nmv_context.comps[0].fp_cdf));
132 memcpy(fc->nmvc_0_hp_cdf, default_nmv_context.comps[0].hp_cdf, sizeof(default_nmv_context.comps[0].hp_cdf));
133 memcpy(fc->nmvc_0_sign_cdf, default_nmv_context.comps[0].sign_cdf, sizeof(default_nmv_context.comps[0].sign_cdf));
134 memcpy(fc->nmvc_1_bits_cdf, default_nmv_context.comps[1].bits_cdf, sizeof(default_nmv_context.comps[1].bits_cdf));
135 memcpy(fc->nmvc_1_class0_cdf, default_nmv_context.comps[1].class0_cdf,
136 sizeof(default_nmv_context.comps[1].class0_cdf));
137 memcpy(fc->nmvc_1_class0_fp_cdf, default_nmv_context.comps[1].class0_fp_cdf,
138 sizeof(default_nmv_context.comps[1].class0_fp_cdf));
139 memcpy(fc->nmvc_1_class0_hp_cdf, default_nmv_context.comps[1].class0_hp_cdf,
140 sizeof(default_nmv_context.comps[1].class0_hp_cdf));
141 memcpy(fc->nmvc_1_classes_cdf, default_nmv_context.comps[1].classes_cdf,
142 sizeof(default_nmv_context.comps[1].classes_cdf));
143 memcpy(fc->nmvc_1_fp_cdf, default_nmv_context.comps[1].fp_cdf, sizeof(default_nmv_context.comps[1].fp_cdf));
144 memcpy(fc->nmvc_1_hp_cdf, default_nmv_context.comps[1].hp_cdf, sizeof(default_nmv_context.comps[1].hp_cdf));
145 memcpy(fc->nmvc_1_sign_cdf, default_nmv_context.comps[1].sign_cdf, sizeof(default_nmv_context.comps[1].sign_cdf));
146 memcpy(fc->ndvc_joints_cdf, default_nmv_context.joints_cdf, sizeof(default_nmv_context.joints_cdf));
147 memcpy(fc->ndvc_0_bits_cdf, default_nmv_context.comps[0].bits_cdf, sizeof(default_nmv_context.comps[0].bits_cdf));
148 memcpy(fc->ndvc_0_class0_cdf, default_nmv_context.comps[0].class0_cdf,
149 sizeof(default_nmv_context.comps[0].class0_cdf));
150 memcpy(fc->ndvc_0_class0_fp_cdf, default_nmv_context.comps[0].class0_fp_cdf,
151 sizeof(default_nmv_context.comps[0].class0_fp_cdf));
152 memcpy(fc->ndvc_0_class0_hp_cdf, default_nmv_context.comps[0].class0_hp_cdf,
153 sizeof(default_nmv_context.comps[0].class0_hp_cdf));
154 memcpy(fc->ndvc_0_classes_cdf, default_nmv_context.comps[0].classes_cdf,
155 sizeof(default_nmv_context.comps[0].classes_cdf));
156 memcpy(fc->ndvc_0_fp_cdf, default_nmv_context.comps[0].fp_cdf, sizeof(default_nmv_context.comps[0].fp_cdf));
157 memcpy(fc->ndvc_0_hp_cdf, default_nmv_context.comps[0].hp_cdf, sizeof(default_nmv_context.comps[0].hp_cdf));
158 memcpy(fc->ndvc_0_sign_cdf, default_nmv_context.comps[0].sign_cdf, sizeof(default_nmv_context.comps[0].sign_cdf));
159 memcpy(fc->ndvc_1_bits_cdf, default_nmv_context.comps[1].bits_cdf, sizeof(default_nmv_context.comps[1].bits_cdf));
160 memcpy(fc->ndvc_1_class0_cdf, default_nmv_context.comps[1].class0_cdf,
161 sizeof(default_nmv_context.comps[1].class0_cdf));
162 memcpy(fc->ndvc_1_class0_fp_cdf, default_nmv_context.comps[1].class0_fp_cdf,
163 sizeof(default_nmv_context.comps[1].class0_fp_cdf));
164 memcpy(fc->ndvc_1_class0_hp_cdf, default_nmv_context.comps[1].class0_hp_cdf,
165 sizeof(default_nmv_context.comps[1].class0_hp_cdf));
166 memcpy(fc->ndvc_1_classes_cdf, default_nmv_context.comps[1].classes_cdf,
167 sizeof(default_nmv_context.comps[1].classes_cdf));
168 memcpy(fc->ndvc_1_fp_cdf, default_nmv_context.comps[1].fp_cdf, sizeof(default_nmv_context.comps[1].fp_cdf));
169 memcpy(fc->ndvc_1_hp_cdf, default_nmv_context.comps[1].hp_cdf, sizeof(default_nmv_context.comps[1].hp_cdf));
170 memcpy(fc->ndvc_1_sign_cdf, default_nmv_context.comps[1].sign_cdf, sizeof(default_nmv_context.comps[1].sign_cdf));
171 }
172
173 static void
ac_vcn_av1_default_coef_probs(void * prob,int index)174 ac_vcn_av1_default_coef_probs(void *prob, int index)
175 {
176 rvcn_av1_frame_context_t *fc = (rvcn_av1_frame_context_t *)prob;
177
178 memcpy(fc->txb_skip_cdf, av1_default_txb_skip_cdfs[index], sizeof(av1_default_txb_skip_cdfs[index]));
179 memcpy(fc->eob_extra_cdf, av1_default_eob_extra_cdfs[index], sizeof(av1_default_eob_extra_cdfs[index]));
180 memcpy(fc->dc_sign_cdf, av1_default_dc_sign_cdfs[index], sizeof(av1_default_dc_sign_cdfs[index]));
181 memcpy(fc->coeff_br_cdf, av1_default_coeff_lps_multi_cdfs[index], sizeof(av1_default_coeff_lps_multi_cdfs[index]));
182 memcpy(fc->coeff_base_cdf, av1_default_coeff_base_multi_cdfs[index],
183 sizeof(av1_default_coeff_base_multi_cdfs[index]));
184 memcpy(fc->coeff_base_eob_cdf, av1_default_coeff_base_eob_multi_cdfs[index],
185 sizeof(av1_default_coeff_base_eob_multi_cdfs[index]));
186 memcpy(fc->eob_flag_cdf16, av1_default_eob_multi16_cdfs[index], sizeof(av1_default_eob_multi16_cdfs[index]));
187 memcpy(fc->eob_flag_cdf32, av1_default_eob_multi32_cdfs[index], sizeof(av1_default_eob_multi32_cdfs[index]));
188 memcpy(fc->eob_flag_cdf64, av1_default_eob_multi64_cdfs[index], sizeof(av1_default_eob_multi64_cdfs[index]));
189 memcpy(fc->eob_flag_cdf128, av1_default_eob_multi128_cdfs[index], sizeof(av1_default_eob_multi128_cdfs[index]));
190 memcpy(fc->eob_flag_cdf256, av1_default_eob_multi256_cdfs[index], sizeof(av1_default_eob_multi256_cdfs[index]));
191 memcpy(fc->eob_flag_cdf512, av1_default_eob_multi512_cdfs[index], sizeof(av1_default_eob_multi512_cdfs[index]));
192 memcpy(fc->eob_flag_cdf1024, av1_default_eob_multi1024_cdfs[index], sizeof(av1_default_eob_multi1024_cdfs[index]));
193 }
194
195 static void
ac_vcn_vcn4_av1_init_mode_probs(void * prob)196 ac_vcn_vcn4_av1_init_mode_probs(void *prob)
197 {
198 rvcn_av1_vcn4_frame_context_t *fc = (rvcn_av1_vcn4_frame_context_t *)prob;
199 int i;
200
201 memcpy(fc->palette_y_size_cdf, default_palette_y_size_cdf, sizeof(default_palette_y_size_cdf));
202 memcpy(fc->palette_uv_size_cdf, default_palette_uv_size_cdf, sizeof(default_palette_uv_size_cdf));
203 memcpy(fc->palette_y_color_index_cdf, default_palette_y_color_index_cdf, sizeof(default_palette_y_color_index_cdf));
204 memcpy(fc->palette_uv_color_index_cdf, default_palette_uv_color_index_cdf,
205 sizeof(default_palette_uv_color_index_cdf));
206 memcpy(fc->kf_y_cdf, default_kf_y_mode_cdf, sizeof(default_kf_y_mode_cdf));
207 memcpy(fc->angle_delta_cdf, default_angle_delta_cdf, sizeof(default_angle_delta_cdf));
208 memcpy(fc->comp_inter_cdf, default_comp_inter_cdf, sizeof(default_comp_inter_cdf));
209 memcpy(fc->comp_ref_type_cdf, default_comp_ref_type_cdf, sizeof(default_comp_ref_type_cdf));
210 memcpy(fc->uni_comp_ref_cdf, default_uni_comp_ref_cdf, sizeof(default_uni_comp_ref_cdf));
211 memcpy(fc->palette_y_mode_cdf, default_palette_y_mode_cdf, sizeof(default_palette_y_mode_cdf));
212 memcpy(fc->palette_uv_mode_cdf, default_palette_uv_mode_cdf, sizeof(default_palette_uv_mode_cdf));
213 memcpy(fc->comp_ref_cdf, default_comp_ref_cdf, sizeof(default_comp_ref_cdf));
214 memcpy(fc->comp_bwdref_cdf, default_comp_bwdref_cdf, sizeof(default_comp_bwdref_cdf));
215 memcpy(fc->single_ref_cdf, default_single_ref_cdf, sizeof(default_single_ref_cdf));
216 memcpy(fc->txfm_partition_cdf, default_txfm_partition_cdf, sizeof(default_txfm_partition_cdf));
217 memcpy(fc->compound_index_cdf, default_compound_idx_cdfs, sizeof(default_compound_idx_cdfs));
218 memcpy(fc->comp_group_idx_cdf, default_comp_group_idx_cdfs, sizeof(default_comp_group_idx_cdfs));
219 memcpy(fc->newmv_cdf, default_newmv_cdf, sizeof(default_newmv_cdf));
220 memcpy(fc->zeromv_cdf, default_zeromv_cdf, sizeof(default_zeromv_cdf));
221 memcpy(fc->refmv_cdf, default_refmv_cdf, sizeof(default_refmv_cdf));
222 memcpy(fc->drl_cdf, default_drl_cdf, sizeof(default_drl_cdf));
223 memcpy(fc->motion_mode_cdf, default_motion_mode_cdf, sizeof(default_motion_mode_cdf));
224 memcpy(fc->obmc_cdf, default_obmc_cdf, sizeof(default_obmc_cdf));
225 memcpy(fc->inter_compound_mode_cdf, default_inter_compound_mode_cdf, sizeof(default_inter_compound_mode_cdf));
226 memcpy(fc->compound_type_cdf, default_compound_type_cdf, sizeof(default_compound_type_cdf));
227 memcpy(fc->wedge_idx_cdf, default_wedge_idx_cdf, sizeof(default_wedge_idx_cdf));
228 memcpy(fc->interintra_cdf, default_interintra_cdf, sizeof(default_interintra_cdf));
229 memcpy(fc->wedge_interintra_cdf, default_wedge_interintra_cdf, sizeof(default_wedge_interintra_cdf));
230 memcpy(fc->interintra_mode_cdf, default_interintra_mode_cdf, sizeof(default_interintra_mode_cdf));
231 memcpy(fc->pred_cdf, default_segment_pred_cdf, sizeof(default_segment_pred_cdf));
232 memcpy(fc->switchable_restore_cdf, default_switchable_restore_cdf, sizeof(default_switchable_restore_cdf));
233 memcpy(fc->wiener_restore_cdf, default_wiener_restore_cdf, sizeof(default_wiener_restore_cdf));
234 memcpy(fc->sgrproj_restore_cdf, default_sgrproj_restore_cdf, sizeof(default_sgrproj_restore_cdf));
235 memcpy(fc->y_mode_cdf, default_if_y_mode_cdf, sizeof(default_if_y_mode_cdf));
236 memcpy(fc->uv_mode_cdf, default_uv_mode_cdf, sizeof(default_uv_mode_cdf));
237 memcpy(fc->switchable_interp_cdf, default_switchable_interp_cdf, sizeof(default_switchable_interp_cdf));
238 memcpy(fc->partition_cdf, default_partition_cdf, sizeof(default_partition_cdf));
239 memcpy(fc->intra_ext_tx_cdf, &default_intra_ext_tx_cdf[1], sizeof(default_intra_ext_tx_cdf[1]) * 2);
240 memcpy(fc->inter_ext_tx_cdf, &default_inter_ext_tx_cdf[1], sizeof(default_inter_ext_tx_cdf[1]) * 3);
241 memcpy(fc->skip_cdfs, default_skip_cdfs, sizeof(default_skip_cdfs));
242 memcpy(fc->intra_inter_cdf, default_intra_inter_cdf, sizeof(default_intra_inter_cdf));
243 memcpy(fc->tree_cdf, default_seg_tree_cdf, sizeof(default_seg_tree_cdf));
244 for (i = 0; i < SPATIAL_PREDICTION_PROBS; ++i)
245 memcpy(fc->spatial_pred_seg_cdf[i], default_spatial_pred_seg_tree_cdf[i],
246 sizeof(default_spatial_pred_seg_tree_cdf[i]));
247 memcpy(fc->tx_size_cdf, default_tx_size_cdf, sizeof(default_tx_size_cdf));
248 memcpy(fc->delta_q_cdf, default_delta_q_cdf, sizeof(default_delta_q_cdf));
249 memcpy(fc->skip_mode_cdfs, default_skip_mode_cdfs, sizeof(default_skip_mode_cdfs));
250 memcpy(fc->delta_lf_cdf, default_delta_lf_cdf, sizeof(default_delta_lf_cdf));
251 memcpy(fc->delta_lf_multi_cdf, default_delta_lf_multi_cdf, sizeof(default_delta_lf_multi_cdf));
252 memcpy(fc->cfl_sign_cdf, default_cfl_sign_cdf, sizeof(default_cfl_sign_cdf));
253 memcpy(fc->cfl_alpha_cdf, default_cfl_alpha_cdf, sizeof(default_cfl_alpha_cdf));
254 memcpy(fc->filter_intra_cdfs, default_filter_intra_cdfs, sizeof(default_filter_intra_cdfs));
255 memcpy(fc->filter_intra_mode_cdf, default_filter_intra_mode_cdf, sizeof(default_filter_intra_mode_cdf));
256 memcpy(fc->intrabc_cdf, default_intrabc_cdf, sizeof(default_intrabc_cdf));
257 }
258
259 static void
ac_vcn_vcn4_av1_init_mv_probs(void * prob)260 ac_vcn_vcn4_av1_init_mv_probs(void *prob)
261 {
262 rvcn_av1_vcn4_frame_context_t *fc = (rvcn_av1_vcn4_frame_context_t *)prob;
263
264 memcpy(fc->nmvc_joints_cdf, default_nmv_context.joints_cdf, sizeof(default_nmv_context.joints_cdf));
265 memcpy(fc->nmvc_0_bits_cdf, default_nmv_context.comps[0].bits_cdf, sizeof(default_nmv_context.comps[0].bits_cdf));
266 memcpy(fc->nmvc_0_class0_cdf, default_nmv_context.comps[0].class0_cdf,
267 sizeof(default_nmv_context.comps[0].class0_cdf));
268 memcpy(fc->nmvc_0_class0_fp_cdf, default_nmv_context.comps[0].class0_fp_cdf,
269 sizeof(default_nmv_context.comps[0].class0_fp_cdf));
270 memcpy(fc->nmvc_0_class0_hp_cdf, default_nmv_context.comps[0].class0_hp_cdf,
271 sizeof(default_nmv_context.comps[0].class0_hp_cdf));
272 memcpy(fc->nmvc_0_classes_cdf, default_nmv_context.comps[0].classes_cdf,
273 sizeof(default_nmv_context.comps[0].classes_cdf));
274 memcpy(fc->nmvc_0_fp_cdf, default_nmv_context.comps[0].fp_cdf, sizeof(default_nmv_context.comps[0].fp_cdf));
275 memcpy(fc->nmvc_0_hp_cdf, default_nmv_context.comps[0].hp_cdf, sizeof(default_nmv_context.comps[0].hp_cdf));
276 memcpy(fc->nmvc_0_sign_cdf, default_nmv_context.comps[0].sign_cdf, sizeof(default_nmv_context.comps[0].sign_cdf));
277 memcpy(fc->nmvc_1_bits_cdf, default_nmv_context.comps[1].bits_cdf, sizeof(default_nmv_context.comps[1].bits_cdf));
278 memcpy(fc->nmvc_1_class0_cdf, default_nmv_context.comps[1].class0_cdf,
279 sizeof(default_nmv_context.comps[1].class0_cdf));
280 memcpy(fc->nmvc_1_class0_fp_cdf, default_nmv_context.comps[1].class0_fp_cdf,
281 sizeof(default_nmv_context.comps[1].class0_fp_cdf));
282 memcpy(fc->nmvc_1_class0_hp_cdf, default_nmv_context.comps[1].class0_hp_cdf,
283 sizeof(default_nmv_context.comps[1].class0_hp_cdf));
284 memcpy(fc->nmvc_1_classes_cdf, default_nmv_context.comps[1].classes_cdf,
285 sizeof(default_nmv_context.comps[1].classes_cdf));
286 memcpy(fc->nmvc_1_fp_cdf, default_nmv_context.comps[1].fp_cdf, sizeof(default_nmv_context.comps[1].fp_cdf));
287 memcpy(fc->nmvc_1_hp_cdf, default_nmv_context.comps[1].hp_cdf, sizeof(default_nmv_context.comps[1].hp_cdf));
288 memcpy(fc->nmvc_1_sign_cdf, default_nmv_context.comps[1].sign_cdf, sizeof(default_nmv_context.comps[1].sign_cdf));
289 memcpy(fc->ndvc_joints_cdf, default_nmv_context.joints_cdf, sizeof(default_nmv_context.joints_cdf));
290 memcpy(fc->ndvc_0_bits_cdf, default_nmv_context.comps[0].bits_cdf, sizeof(default_nmv_context.comps[0].bits_cdf));
291 memcpy(fc->ndvc_0_class0_cdf, default_nmv_context.comps[0].class0_cdf,
292 sizeof(default_nmv_context.comps[0].class0_cdf));
293 memcpy(fc->ndvc_0_class0_fp_cdf, default_nmv_context.comps[0].class0_fp_cdf,
294 sizeof(default_nmv_context.comps[0].class0_fp_cdf));
295 memcpy(fc->ndvc_0_class0_hp_cdf, default_nmv_context.comps[0].class0_hp_cdf,
296 sizeof(default_nmv_context.comps[0].class0_hp_cdf));
297 memcpy(fc->ndvc_0_classes_cdf, default_nmv_context.comps[0].classes_cdf,
298 sizeof(default_nmv_context.comps[0].classes_cdf));
299 memcpy(fc->ndvc_0_fp_cdf, default_nmv_context.comps[0].fp_cdf, sizeof(default_nmv_context.comps[0].fp_cdf));
300 memcpy(fc->ndvc_0_hp_cdf, default_nmv_context.comps[0].hp_cdf, sizeof(default_nmv_context.comps[0].hp_cdf));
301 memcpy(fc->ndvc_0_sign_cdf, default_nmv_context.comps[0].sign_cdf, sizeof(default_nmv_context.comps[0].sign_cdf));
302 memcpy(fc->ndvc_1_bits_cdf, default_nmv_context.comps[1].bits_cdf, sizeof(default_nmv_context.comps[1].bits_cdf));
303 memcpy(fc->ndvc_1_class0_cdf, default_nmv_context.comps[1].class0_cdf,
304 sizeof(default_nmv_context.comps[1].class0_cdf));
305 memcpy(fc->ndvc_1_class0_fp_cdf, default_nmv_context.comps[1].class0_fp_cdf,
306 sizeof(default_nmv_context.comps[1].class0_fp_cdf));
307 memcpy(fc->ndvc_1_class0_hp_cdf, default_nmv_context.comps[1].class0_hp_cdf,
308 sizeof(default_nmv_context.comps[1].class0_hp_cdf));
309 memcpy(fc->ndvc_1_classes_cdf, default_nmv_context.comps[1].classes_cdf,
310 sizeof(default_nmv_context.comps[1].classes_cdf));
311 memcpy(fc->ndvc_1_fp_cdf, default_nmv_context.comps[1].fp_cdf, sizeof(default_nmv_context.comps[1].fp_cdf));
312 memcpy(fc->ndvc_1_hp_cdf, default_nmv_context.comps[1].hp_cdf, sizeof(default_nmv_context.comps[1].hp_cdf));
313 memcpy(fc->ndvc_1_sign_cdf, default_nmv_context.comps[1].sign_cdf, sizeof(default_nmv_context.comps[1].sign_cdf));
314 }
315
316 static void
ac_vcn_vcn4_av1_default_coef_probs(void * prob,int index)317 ac_vcn_vcn4_av1_default_coef_probs(void *prob, int index)
318 {
319 rvcn_av1_vcn4_frame_context_t *fc = (rvcn_av1_vcn4_frame_context_t *)prob;
320 char *p;
321 int i, j;
322 unsigned size;
323
324 memcpy(fc->txb_skip_cdf, av1_default_txb_skip_cdfs[index], sizeof(av1_default_txb_skip_cdfs[index]));
325
326 p = (char *)fc->eob_extra_cdf;
327 size = sizeof(av1_default_eob_extra_cdfs[0][0][0][0]) * EOB_COEF_CONTEXTS_VCN4;
328 for (i = 0; i < AV1_TX_SIZES; i++) {
329 for (j = 0; j < AV1_PLANE_TYPES; j++) {
330 memcpy(p, &av1_default_eob_extra_cdfs[index][i][j][3], size);
331 p += size;
332 }
333 }
334
335 memcpy(fc->dc_sign_cdf, av1_default_dc_sign_cdfs[index], sizeof(av1_default_dc_sign_cdfs[index]));
336 memcpy(fc->coeff_br_cdf, av1_default_coeff_lps_multi_cdfs[index], sizeof(av1_default_coeff_lps_multi_cdfs[index]));
337 memcpy(fc->coeff_base_cdf, av1_default_coeff_base_multi_cdfs[index],
338 sizeof(av1_default_coeff_base_multi_cdfs[index]));
339 memcpy(fc->coeff_base_eob_cdf, av1_default_coeff_base_eob_multi_cdfs[index],
340 sizeof(av1_default_coeff_base_eob_multi_cdfs[index]));
341 memcpy(fc->eob_flag_cdf16, av1_default_eob_multi16_cdfs[index], sizeof(av1_default_eob_multi16_cdfs[index]));
342 memcpy(fc->eob_flag_cdf32, av1_default_eob_multi32_cdfs[index], sizeof(av1_default_eob_multi32_cdfs[index]));
343 memcpy(fc->eob_flag_cdf64, av1_default_eob_multi64_cdfs[index], sizeof(av1_default_eob_multi64_cdfs[index]));
344 memcpy(fc->eob_flag_cdf128, av1_default_eob_multi128_cdfs[index], sizeof(av1_default_eob_multi128_cdfs[index]));
345 memcpy(fc->eob_flag_cdf256, av1_default_eob_multi256_cdfs[index], sizeof(av1_default_eob_multi256_cdfs[index]));
346 memcpy(fc->eob_flag_cdf512, av1_default_eob_multi512_cdfs[index], sizeof(av1_default_eob_multi512_cdfs[index]));
347 memcpy(fc->eob_flag_cdf1024, av1_default_eob_multi1024_cdfs[index], sizeof(av1_default_eob_multi1024_cdfs[index]));
348 }
349
350 void
ac_vcn_av1_init_probs(unsigned av1_version,uint8_t * prob)351 ac_vcn_av1_init_probs(unsigned av1_version, uint8_t *prob)
352 {
353 unsigned frame_ctxt_size = ac_vcn_dec_frame_ctx_size_av1(av1_version);
354 if (av1_version == RDECODE_AV1_VER_0) {
355 for (unsigned i = 0; i < 4; ++i) {
356 ac_vcn_av1_init_mode_probs((void *)(prob + i * frame_ctxt_size));
357 ac_vcn_av1_init_mv_probs((void *)(prob + i * frame_ctxt_size));
358 ac_vcn_av1_default_coef_probs((void *)(prob + i * frame_ctxt_size), i);
359 }
360 } else {
361 for (unsigned i = 0; i < 4; ++i) {
362 ac_vcn_vcn4_av1_init_mode_probs((void *)(prob + i * frame_ctxt_size));
363 ac_vcn_vcn4_av1_init_mv_probs((void *)(prob + i * frame_ctxt_size));
364 ac_vcn_vcn4_av1_default_coef_probs((void *)(prob + i * frame_ctxt_size), i);
365 }
366 }
367
368 }
369
370 #define LUMA_BLOCK_SIZE_Y 73
371 #define LUMA_BLOCK_SIZE_X 82
372 #define CHROMA_BLOCK_SIZE_Y 38
373 #define CHROMA_BLOCK_SIZE_X 44
374
375 static int32_t
radv_vcn_av1_film_grain_random_number(unsigned short * seed,int32_t bits)376 radv_vcn_av1_film_grain_random_number(unsigned short *seed, int32_t bits)
377 {
378 unsigned short bit;
379 unsigned short value = *seed;
380
381 bit = ((value >> 0) ^ (value >> 1) ^ (value >> 3) ^ (value >> 12)) & 1;
382 value = (value >> 1) | (bit << 15);
383 *seed = value;
384
385 return (value >> (16 - bits)) & ((1 << bits) - 1);
386 }
387
388 static void
radv_vcn_av1_film_grain_init_scaling(uint8_t scaling_points[][2],uint8_t num,short scaling_lut[])389 radv_vcn_av1_film_grain_init_scaling(uint8_t scaling_points[][2], uint8_t num, short scaling_lut[])
390 {
391 int32_t i, x, delta_x, delta_y;
392 int64_t delta;
393
394 if (num == 0)
395 return;
396
397 for (i = 0; i < scaling_points[0][0]; i++)
398 scaling_lut[i] = scaling_points[0][1];
399
400 for (i = 0; i < num - 1; i++) {
401 delta_y = scaling_points[i + 1][1] - scaling_points[i][1];
402 delta_x = scaling_points[i + 1][0] - scaling_points[i][0];
403
404 delta = delta_y * ((65536 + (delta_x >> 1)) / delta_x);
405
406 for (x = 0; x < delta_x; x++)
407 scaling_lut[scaling_points[i][0] + x] = (short)(scaling_points[i][1] + (int32_t)((x * delta + 32768) >> 16));
408 }
409
410 for (i = scaling_points[num - 1][0]; i < 256; i++)
411 scaling_lut[i] = scaling_points[num - 1][1];
412 }
413
414 void
ac_vcn_av1_init_film_grain_buffer(rvcn_dec_film_grain_params_t * fg_params,rvcn_dec_av1_fg_init_buf_t * fg_buf)415 ac_vcn_av1_init_film_grain_buffer(rvcn_dec_film_grain_params_t *fg_params, rvcn_dec_av1_fg_init_buf_t *fg_buf)
416 {
417 const int32_t luma_block_size_y = LUMA_BLOCK_SIZE_Y;
418 const int32_t luma_block_size_x = LUMA_BLOCK_SIZE_X;
419 const int32_t chroma_block_size_y = CHROMA_BLOCK_SIZE_Y;
420 const int32_t chroma_block_size_x = CHROMA_BLOCK_SIZE_X;
421 const int32_t gauss_bits = 11;
422 int32_t filt_luma_grain_block[LUMA_BLOCK_SIZE_Y][LUMA_BLOCK_SIZE_X];
423 int32_t filt_cb_grain_block[CHROMA_BLOCK_SIZE_Y][CHROMA_BLOCK_SIZE_X];
424 int32_t filt_cr_grain_block[CHROMA_BLOCK_SIZE_Y][CHROMA_BLOCK_SIZE_X];
425 int32_t chroma_subsamp_y = 1;
426 int32_t chroma_subsamp_x = 1;
427 unsigned short seed = fg_params->random_seed;
428 int32_t ar_coeff_lag = fg_params->ar_coeff_lag;
429 int32_t bit_depth = fg_params->bit_depth_minus_8 + 8;
430 short grain_center = 128 << (bit_depth - 8);
431 short grain_min = 0 - grain_center;
432 short grain_max = (256 << (bit_depth - 8)) - 1 - grain_center;
433 int32_t shift = 12 - bit_depth + fg_params->grain_scale_shift;
434 short luma_grain_block_tmp[64][80];
435 short cb_grain_block_tmp[32][40];
436 short cr_grain_block_tmp[32][40];
437 short *align_ptr, *align_ptr0, *align_ptr1;
438 int32_t x, y, g, i, j, c, c0, c1, delta_row, delta_col;
439 int32_t s, s0, s1, pos, r;
440
441 /* generate luma grain block */
442 memset(filt_luma_grain_block, 0, sizeof(filt_luma_grain_block));
443 for (y = 0; y < luma_block_size_y; y++) {
444 for (x = 0; x < luma_block_size_x; x++) {
445 g = 0;
446 if (fg_params->num_y_points > 0) {
447 r = radv_vcn_av1_film_grain_random_number(&seed, gauss_bits);
448 g = gaussian_sequence[CLAMP(r, 0, 2048 - 1)];
449 }
450 filt_luma_grain_block[y][x] = ROUND_POWER_OF_TWO(g, shift);
451 }
452 }
453
454 for (y = 3; y < luma_block_size_y; y++) {
455 for (x = 3; x < luma_block_size_x - 3; x++) {
456 s = 0;
457 pos = 0;
458 for (delta_row = -ar_coeff_lag; delta_row <= 0; delta_row++) {
459 for (delta_col = -ar_coeff_lag; delta_col <= ar_coeff_lag; delta_col++) {
460 if (delta_row == 0 && delta_col == 0)
461 break;
462 c = fg_params->ar_coeffs_y[pos];
463 s += filt_luma_grain_block[y + delta_row][x + delta_col] * c;
464 pos++;
465 }
466 }
467 filt_luma_grain_block[y][x] = AV1_CLAMP(
468 filt_luma_grain_block[y][x] + ROUND_POWER_OF_TWO(s, fg_params->ar_coeff_shift), grain_min, grain_max);
469 }
470 }
471
472 /* generate chroma grain block */
473 memset(filt_cb_grain_block, 0, sizeof(filt_cb_grain_block));
474 shift = 12 - bit_depth + fg_params->grain_scale_shift;
475 seed = fg_params->random_seed ^ 0xb524;
476 for (y = 0; y < chroma_block_size_y; y++) {
477 for (x = 0; x < chroma_block_size_x; x++) {
478 g = 0;
479 if (fg_params->num_cb_points || fg_params->chroma_scaling_from_luma) {
480 r = radv_vcn_av1_film_grain_random_number(&seed, gauss_bits);
481 g = gaussian_sequence[CLAMP(r, 0, 2048 - 1)];
482 }
483 filt_cb_grain_block[y][x] = ROUND_POWER_OF_TWO(g, shift);
484 }
485 }
486
487 memset(filt_cr_grain_block, 0, sizeof(filt_cr_grain_block));
488 seed = fg_params->random_seed ^ 0x49d8;
489 for (y = 0; y < chroma_block_size_y; y++) {
490 for (x = 0; x < chroma_block_size_x; x++) {
491 g = 0;
492 if (fg_params->num_cr_points || fg_params->chroma_scaling_from_luma) {
493 r = radv_vcn_av1_film_grain_random_number(&seed, gauss_bits);
494 g = gaussian_sequence[CLAMP(r, 0, 2048 - 1)];
495 }
496 filt_cr_grain_block[y][x] = ROUND_POWER_OF_TWO(g, shift);
497 }
498 }
499
500 for (y = 3; y < chroma_block_size_y; y++) {
501 for (x = 3; x < chroma_block_size_x - 3; x++) {
502 s0 = 0, s1 = 0, pos = 0;
503 for (delta_row = -ar_coeff_lag; delta_row <= 0; delta_row++) {
504 for (delta_col = -ar_coeff_lag; delta_col <= ar_coeff_lag; delta_col++) {
505 c0 = fg_params->ar_coeffs_cb[pos];
506 c1 = fg_params->ar_coeffs_cr[pos];
507 if (delta_row == 0 && delta_col == 0) {
508 if (fg_params->num_y_points > 0) {
509 int luma = 0;
510 int luma_x = ((x - 3) << chroma_subsamp_x) + 3;
511 int luma_y = ((y - 3) << chroma_subsamp_y) + 3;
512 for (i = 0; i <= chroma_subsamp_y; i++)
513 for (j = 0; j <= chroma_subsamp_x; j++)
514 luma += filt_luma_grain_block[luma_y + i][luma_x + j];
515
516 luma = ROUND_POWER_OF_TWO(luma, chroma_subsamp_x + chroma_subsamp_y);
517 s0 += luma * c0;
518 s1 += luma * c1;
519 }
520 break;
521 }
522 s0 += filt_cb_grain_block[y + delta_row][x + delta_col] * c0;
523 s1 += filt_cr_grain_block[y + delta_row][x + delta_col] * c1;
524 pos++;
525 }
526 }
527 filt_cb_grain_block[y][x] = AV1_CLAMP(
528 filt_cb_grain_block[y][x] + ROUND_POWER_OF_TWO(s0, fg_params->ar_coeff_shift), grain_min, grain_max);
529 filt_cr_grain_block[y][x] = AV1_CLAMP(
530 filt_cr_grain_block[y][x] + ROUND_POWER_OF_TWO(s1, fg_params->ar_coeff_shift), grain_min, grain_max);
531 }
532 }
533
534 for (i = 9; i < luma_block_size_y; i++)
535 for (j = 9; j < luma_block_size_x; j++)
536 luma_grain_block_tmp[i - 9][j - 9] = filt_luma_grain_block[i][j];
537
538 for (i = 6; i < chroma_block_size_y; i++)
539 for (j = 6; j < chroma_block_size_x; j++) {
540 cb_grain_block_tmp[i - 6][j - 6] = filt_cb_grain_block[i][j];
541 cr_grain_block_tmp[i - 6][j - 6] = filt_cr_grain_block[i][j];
542 }
543
544 align_ptr = &fg_buf->luma_grain_block[0][0];
545 for (i = 0; i < 64; i++) {
546 for (j = 0; j < 80; j++)
547 *align_ptr++ = luma_grain_block_tmp[i][j];
548
549 if (((i + 1) % 4) == 0)
550 align_ptr += 64;
551 }
552
553 align_ptr0 = &fg_buf->cb_grain_block[0][0];
554 align_ptr1 = &fg_buf->cr_grain_block[0][0];
555 for (i = 0; i < 32; i++) {
556 for (j = 0; j < 40; j++) {
557 *align_ptr0++ = cb_grain_block_tmp[i][j];
558 *align_ptr1++ = cr_grain_block_tmp[i][j];
559 }
560 if (((i + 1) % 8) == 0) {
561 align_ptr0 += 64;
562 align_ptr1 += 64;
563 }
564 }
565
566 memset(fg_buf->scaling_lut_y, 0, sizeof(fg_buf->scaling_lut_y));
567 radv_vcn_av1_film_grain_init_scaling(fg_params->scaling_points_y, fg_params->num_y_points, fg_buf->scaling_lut_y);
568 if (fg_params->chroma_scaling_from_luma) {
569 memcpy(fg_buf->scaling_lut_cb, fg_buf->scaling_lut_y, sizeof(fg_buf->scaling_lut_y));
570 memcpy(fg_buf->scaling_lut_cr, fg_buf->scaling_lut_y, sizeof(fg_buf->scaling_lut_y));
571 } else {
572 memset(fg_buf->scaling_lut_cb, 0, sizeof(fg_buf->scaling_lut_cb));
573 memset(fg_buf->scaling_lut_cr, 0, sizeof(fg_buf->scaling_lut_cr));
574 radv_vcn_av1_film_grain_init_scaling(fg_params->scaling_points_cb, fg_params->num_cb_points,
575 fg_buf->scaling_lut_cb);
576 radv_vcn_av1_film_grain_init_scaling(fg_params->scaling_points_cr, fg_params->num_cr_points,
577 fg_buf->scaling_lut_cr);
578 }
579 }
580