xref: /aosp_15_r20/external/libaom/av1/encoder/model_rd.h (revision 77c1e3ccc04c968bd2bc212e87364f250e820521)
1 /*
2  * Copyright (c) 2020, Alliance for Open Media. All rights reserved.
3  *
4  * This source code is subject to the terms of the BSD 2 Clause License and
5  * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6  * was not distributed with this source code in the LICENSE file, you can
7  * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8  * Media Patent License 1.0 was not distributed with this source code in the
9  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10  */
11 
12 #ifndef AOM_AV1_ENCODER_MODEL_RD_H_
13 #define AOM_AV1_ENCODER_MODEL_RD_H_
14 
15 #include "aom/aom_integer.h"
16 #include "av1/encoder/block.h"
17 #include "av1/encoder/encoder.h"
18 #include "av1/encoder/pustats.h"
19 #include "av1/encoder/rdopt_utils.h"
20 #include "config/aom_dsp_rtcd.h"
21 
22 #ifdef __cplusplus
23 extern "C" {
24 #endif
25 
26 // 0: Legacy model
27 // 1: Curve fit model
28 // 2: Surface fit model
29 // 3: DNN regression model
30 // 4: Full rd model
31 #define MODELRD_TYPE_INTERP_FILTER 1
32 #define MODELRD_TYPE_TX_SEARCH_PRUNE 1
33 #define MODELRD_TYPE_MASKED_COMPOUND 1
34 #define MODELRD_TYPE_INTERINTRA 1
35 #define MODELRD_TYPE_INTRA 1
36 #define MODELRD_TYPE_MOTION_MODE_RD 1
37 
38 typedef void (*model_rd_for_sb_type)(
39     const AV1_COMP *const cpi, BLOCK_SIZE bsize, MACROBLOCK *x, MACROBLOCKD *xd,
40     int plane_from, int plane_to, int *out_rate_sum, int64_t *out_dist_sum,
41     uint8_t *skip_txfm_sb, int64_t *skip_sse_sb, int *plane_rate,
42     int64_t *plane_sse, int64_t *plane_dist);
43 typedef void (*model_rd_from_sse_type)(const AV1_COMP *const cpi,
44                                        const MACROBLOCK *const x,
45                                        BLOCK_SIZE plane_bsize, int plane,
46                                        int64_t sse, int num_samples, int *rate,
47                                        int64_t *dist);
48 
calculate_sse(MACROBLOCKD * const xd,const struct macroblock_plane * p,struct macroblockd_plane * pd,const int bw,const int bh)49 static int64_t calculate_sse(MACROBLOCKD *const xd,
50                              const struct macroblock_plane *p,
51                              struct macroblockd_plane *pd, const int bw,
52                              const int bh) {
53   int64_t sse = 0;
54   const int shift = xd->bd - 8;
55 #if CONFIG_AV1_HIGHBITDEPTH
56   if (is_cur_buf_hbd(xd)) {
57     sse = aom_highbd_sse(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
58                          bw, bh);
59   } else {
60     sse =
61         aom_sse(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, bw, bh);
62   }
63 #else
64   sse = aom_sse(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, bw, bh);
65 #endif
66   sse = ROUND_POWER_OF_TWO(sse, shift * 2);
67   return sse;
68 }
69 
compute_sse_plane(MACROBLOCK * x,MACROBLOCKD * xd,int plane,const BLOCK_SIZE bsize)70 static inline int64_t compute_sse_plane(MACROBLOCK *x, MACROBLOCKD *xd,
71                                         int plane, const BLOCK_SIZE bsize) {
72   struct macroblockd_plane *const pd = &xd->plane[plane];
73   const BLOCK_SIZE plane_bsize =
74       get_plane_block_size(bsize, pd->subsampling_x, pd->subsampling_y);
75   int bw, bh;
76   const struct macroblock_plane *const p = &x->plane[plane];
77   get_txb_dimensions(xd, plane, plane_bsize, 0, 0, plane_bsize, NULL, NULL, &bw,
78                      &bh);
79 
80   int64_t sse = calculate_sse(xd, p, pd, bw, bh);
81 
82   return sse;
83 }
84 
model_rd_from_sse(const AV1_COMP * const cpi,const MACROBLOCK * const x,BLOCK_SIZE plane_bsize,int plane,int64_t sse,int num_samples,int * rate,int64_t * dist)85 static inline void model_rd_from_sse(const AV1_COMP *const cpi,
86                                      const MACROBLOCK *const x,
87                                      BLOCK_SIZE plane_bsize, int plane,
88                                      int64_t sse, int num_samples, int *rate,
89                                      int64_t *dist) {
90   (void)num_samples;
91   const MACROBLOCKD *const xd = &x->e_mbd;
92   const struct macroblock_plane *const p = &x->plane[plane];
93   const int dequant_shift = (is_cur_buf_hbd(xd)) ? xd->bd - 5 : 3;
94 
95   // Fast approximate the modelling function.
96   if (cpi->sf.rd_sf.simple_model_rd_from_var) {
97     const int64_t square_error = sse;
98     int quantizer = p->dequant_QTX[1] >> dequant_shift;
99     if (quantizer < 120)
100       *rate = (int)AOMMIN(
101           (square_error * (280 - quantizer)) >> (16 - AV1_PROB_COST_SHIFT),
102           INT_MAX);
103     else
104       *rate = 0;
105     assert(*rate >= 0);
106     *dist = (square_error * quantizer) >> 8;
107   } else {
108     av1_model_rd_from_var_lapndz(sse, num_pels_log2_lookup[plane_bsize],
109                                  p->dequant_QTX[1] >> dequant_shift, rate,
110                                  dist);
111   }
112   *dist <<= 4;
113 }
114 
115 // Fits a curve for rate and distortion using as feature:
116 // log2(sse_norm/qstep^2)
model_rd_with_curvfit(const AV1_COMP * const cpi,const MACROBLOCK * const x,BLOCK_SIZE plane_bsize,int plane,int64_t sse,int num_samples,int * rate,int64_t * dist)117 static inline void model_rd_with_curvfit(const AV1_COMP *const cpi,
118                                          const MACROBLOCK *const x,
119                                          BLOCK_SIZE plane_bsize, int plane,
120                                          int64_t sse, int num_samples,
121                                          int *rate, int64_t *dist) {
122   (void)cpi;
123   (void)plane_bsize;
124   const MACROBLOCKD *const xd = &x->e_mbd;
125   const struct macroblock_plane *const p = &x->plane[plane];
126   const int dequant_shift = (is_cur_buf_hbd(xd)) ? xd->bd - 5 : 3;
127   const int qstep = AOMMAX(p->dequant_QTX[1] >> dequant_shift, 1);
128 
129   if (sse == 0) {
130     if (rate) *rate = 0;
131     if (dist) *dist = 0;
132     return;
133   }
134   const double sse_norm = (double)sse / num_samples;
135   const double qstepsqr = (double)qstep * qstep;
136   const double xqr = log2(sse_norm / qstepsqr);
137   double rate_f, dist_by_sse_norm_f;
138   av1_model_rd_curvfit(plane_bsize, sse_norm, xqr, &rate_f,
139                        &dist_by_sse_norm_f);
140 
141   const double dist_f = dist_by_sse_norm_f * sse_norm;
142   int rate_i = (int)(AOMMAX(0.0, rate_f * num_samples) + 0.5);
143   int64_t dist_i = (int64_t)(AOMMAX(0.0, dist_f * num_samples) + 0.5);
144 
145   // Check if skip is better
146   if (rate_i == 0) {
147     dist_i = sse << 4;
148   } else if (RDCOST(x->rdmult, rate_i, dist_i) >=
149              RDCOST(x->rdmult, 0, sse << 4)) {
150     rate_i = 0;
151     dist_i = sse << 4;
152   }
153 
154   if (rate) *rate = rate_i;
155   if (dist) *dist = dist_i;
156 }
157 
model_rd_for_sb(const AV1_COMP * const cpi,BLOCK_SIZE bsize,MACROBLOCK * x,MACROBLOCKD * xd,int plane_from,int plane_to,int * out_rate_sum,int64_t * out_dist_sum,uint8_t * skip_txfm_sb,int64_t * skip_sse_sb,int * plane_rate,int64_t * plane_sse,int64_t * plane_dist)158 static inline void model_rd_for_sb(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
159                                    MACROBLOCK *x, MACROBLOCKD *xd,
160                                    int plane_from, int plane_to,
161                                    int *out_rate_sum, int64_t *out_dist_sum,
162                                    uint8_t *skip_txfm_sb, int64_t *skip_sse_sb,
163                                    int *plane_rate, int64_t *plane_sse,
164                                    int64_t *plane_dist) {
165   // Note our transform coeffs are 8 times an orthogonal transform.
166   // Hence quantizer step is also 8 times. To get effective quantizer
167   // we need to divide by 8 before sending to modeling function.
168   int plane;
169   const int ref = xd->mi[0]->ref_frame[0];
170 
171   int64_t rate_sum = 0;
172   int64_t dist_sum = 0;
173   int64_t total_sse = 0;
174 
175   assert(bsize < BLOCK_SIZES_ALL);
176 
177   for (plane = plane_from; plane <= plane_to; ++plane) {
178     if (plane && !xd->is_chroma_ref) break;
179     struct macroblock_plane *const p = &x->plane[plane];
180     struct macroblockd_plane *const pd = &xd->plane[plane];
181     const BLOCK_SIZE plane_bsize =
182         get_plane_block_size(bsize, pd->subsampling_x, pd->subsampling_y);
183     assert(plane_bsize < BLOCK_SIZES_ALL);
184     const int bw = block_size_wide[plane_bsize];
185     const int bh = block_size_high[plane_bsize];
186     int64_t sse;
187     int rate;
188     int64_t dist;
189 
190     sse = calculate_sse(xd, p, pd, bw, bh);
191 
192     model_rd_from_sse(cpi, x, plane_bsize, plane, sse, bw * bh, &rate, &dist);
193 
194     if (plane == 0) x->pred_sse[ref] = (unsigned int)AOMMIN(sse, UINT_MAX);
195 
196     total_sse += sse;
197     rate_sum += rate;
198     dist_sum += dist;
199     if (plane_rate) plane_rate[plane] = rate;
200     if (plane_sse) plane_sse[plane] = sse;
201     if (plane_dist) plane_dist[plane] = dist;
202     assert(rate_sum >= 0);
203   }
204 
205   if (skip_txfm_sb) *skip_txfm_sb = total_sse == 0;
206   if (skip_sse_sb) *skip_sse_sb = total_sse << 4;
207   rate_sum = AOMMIN(rate_sum, INT_MAX);
208   *out_rate_sum = (int)rate_sum;
209   *out_dist_sum = dist_sum;
210 }
211 
model_rd_for_sb_with_curvfit(const AV1_COMP * const cpi,BLOCK_SIZE bsize,MACROBLOCK * x,MACROBLOCKD * xd,int plane_from,int plane_to,int * out_rate_sum,int64_t * out_dist_sum,uint8_t * skip_txfm_sb,int64_t * skip_sse_sb,int * plane_rate,int64_t * plane_sse,int64_t * plane_dist)212 static inline void model_rd_for_sb_with_curvfit(
213     const AV1_COMP *const cpi, BLOCK_SIZE bsize, MACROBLOCK *x, MACROBLOCKD *xd,
214     int plane_from, int plane_to, int *out_rate_sum, int64_t *out_dist_sum,
215     uint8_t *skip_txfm_sb, int64_t *skip_sse_sb, int *plane_rate,
216     int64_t *plane_sse, int64_t *plane_dist) {
217   // Note our transform coeffs are 8 times an orthogonal transform.
218   // Hence quantizer step is also 8 times. To get effective quantizer
219   // we need to divide by 8 before sending to modeling function.
220   const int ref = xd->mi[0]->ref_frame[0];
221 
222   int64_t rate_sum = 0;
223   int64_t dist_sum = 0;
224   int64_t total_sse = 0;
225 
226   for (int plane = plane_from; plane <= plane_to; ++plane) {
227     if (plane && !xd->is_chroma_ref) break;
228     struct macroblockd_plane *const pd = &xd->plane[plane];
229     const BLOCK_SIZE plane_bsize =
230         get_plane_block_size(bsize, pd->subsampling_x, pd->subsampling_y);
231     int64_t dist, sse;
232     int rate;
233     int bw, bh;
234     const struct macroblock_plane *const p = &x->plane[plane];
235     get_txb_dimensions(xd, plane, plane_bsize, 0, 0, plane_bsize, NULL, NULL,
236                        &bw, &bh);
237 
238     sse = calculate_sse(xd, p, pd, bw, bh);
239     model_rd_with_curvfit(cpi, x, plane_bsize, plane, sse, bw * bh, &rate,
240                           &dist);
241 
242     if (plane == 0) x->pred_sse[ref] = (unsigned int)AOMMIN(sse, UINT_MAX);
243 
244     total_sse += sse;
245     rate_sum += rate;
246     dist_sum += dist;
247 
248     if (plane_rate) plane_rate[plane] = rate;
249     if (plane_sse) plane_sse[plane] = sse;
250     if (plane_dist) plane_dist[plane] = dist;
251   }
252 
253   if (skip_txfm_sb) *skip_txfm_sb = rate_sum == 0;
254   if (skip_sse_sb) *skip_sse_sb = total_sse << 4;
255   *out_rate_sum = (int)rate_sum;
256   *out_dist_sum = dist_sum;
257 }
258 
259 enum { MODELRD_LEGACY, MODELRD_CURVFIT, MODELRD_TYPES } UENUM1BYTE(ModelRdType);
260 
261 static const model_rd_for_sb_type model_rd_sb_fn[MODELRD_TYPES] = {
262   model_rd_for_sb, model_rd_for_sb_with_curvfit
263 };
264 
265 static const model_rd_from_sse_type model_rd_sse_fn[MODELRD_TYPES] = {
266   model_rd_from_sse, model_rd_with_curvfit
267 };
268 
269 #ifdef __cplusplus
270 }  // extern "C"
271 #endif
272 #endif  // AOM_AV1_ENCODER_MODEL_RD_H_
273