AOMedia AV1 Codec
intra_mode_search_utils.h
Go to the documentation of this file.
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
17#ifndef AOM_AV1_ENCODER_INTRA_MODE_SEARCH_UTILS_H_
18#define AOM_AV1_ENCODER_INTRA_MODE_SEARCH_UTILS_H_
19
20#include "av1/common/enums.h"
21#include "av1/common/pred_common.h"
22#include "av1/common/reconintra.h"
23
24#include "av1/encoder/encoder.h"
25#include "av1/encoder/encodeframe.h"
26#include "av1/encoder/model_rd.h"
27#include "av1/encoder/palette.h"
28#include "av1/encoder/hybrid_fwd_txfm.h"
29
30#ifdef __cplusplus
31extern "C" {
32#endif
33
35#define BINS 32
36static const float av1_intra_hog_model_bias[DIRECTIONAL_MODES] = {
37 0.450578f, 0.695518f, -0.717944f, -0.639894f,
38 -0.602019f, -0.453454f, 0.055857f, -0.465480f,
39};
40
41static const float av1_intra_hog_model_weights[BINS * DIRECTIONAL_MODES] = {
42 -3.076402f, -3.757063f, -3.275266f, -3.180665f, -3.452105f, -3.216593f,
43 -2.871212f, -3.134296f, -1.822324f, -2.401411f, -1.541016f, -1.195322f,
44 -0.434156f, 0.322868f, 2.260546f, 3.368715f, 3.989290f, 3.308487f,
45 2.277893f, 0.923793f, 0.026412f, -0.385174f, -0.718622f, -1.408867f,
46 -1.050558f, -2.323941f, -2.225827f, -2.585453f, -3.054283f, -2.875087f,
47 -2.985709f, -3.447155f, 3.758139f, 3.204353f, 2.170998f, 0.826587f,
48 -0.269665f, -0.702068f, -1.085776f, -2.175249f, -1.623180f, -2.975142f,
49 -2.779629f, -3.190799f, -3.521900f, -3.375480f, -3.319355f, -3.897389f,
50 -3.172334f, -3.594528f, -2.879132f, -2.547777f, -2.921023f, -2.281844f,
51 -1.818988f, -2.041771f, -0.618268f, -1.396458f, -0.567153f, -0.285868f,
52 -0.088058f, 0.753494f, 2.092413f, 3.215266f, -3.300277f, -2.748658f,
53 -2.315784f, -2.423671f, -2.257283f, -2.269583f, -2.196660f, -2.301076f,
54 -2.646516f, -2.271319f, -2.254366f, -2.300102f, -2.217960f, -2.473300f,
55 -2.116866f, -2.528246f, -3.314712f, -1.701010f, -0.589040f, -0.088077f,
56 0.813112f, 1.702213f, 2.653045f, 3.351749f, 3.243554f, 3.199409f,
57 2.437856f, 1.468854f, 0.533039f, -0.099065f, -0.622643f, -2.200732f,
58 -4.228861f, -2.875263f, -1.273956f, -0.433280f, 0.803771f, 1.975043f,
59 3.179528f, 3.939064f, 3.454379f, 3.689386f, 3.116411f, 1.970991f,
60 0.798406f, -0.628514f, -1.252546f, -2.825176f, -4.090178f, -3.777448f,
61 -3.227314f, -3.479403f, -3.320569f, -3.159372f, -2.729202f, -2.722341f,
62 -3.054913f, -2.742923f, -2.612703f, -2.662632f, -2.907314f, -3.117794f,
63 -3.102660f, -3.970972f, -4.891357f, -3.935582f, -3.347758f, -2.721924f,
64 -2.219011f, -1.702391f, -0.866529f, -0.153743f, 0.107733f, 1.416882f,
65 2.572884f, 3.607755f, 3.974820f, 3.997783f, 2.970459f, 0.791687f,
66 -1.478921f, -1.228154f, -1.216955f, -1.765932f, -1.951003f, -1.985301f,
67 -1.975881f, -1.985593f, -2.422371f, -2.419978f, -2.531288f, -2.951853f,
68 -3.071380f, -3.277027f, -3.373539f, -4.462010f, -0.967888f, 0.805524f,
69 2.794130f, 3.685984f, 3.745195f, 3.252444f, 2.316108f, 1.399146f,
70 -0.136519f, -0.162811f, -1.004357f, -1.667911f, -1.964662f, -2.937579f,
71 -3.019533f, -3.942766f, -5.102767f, -3.882073f, -3.532027f, -3.451956f,
72 -2.944015f, -2.643064f, -2.529872f, -2.077290f, -2.809965f, -1.803734f,
73 -1.783593f, -1.662585f, -1.415484f, -1.392673f, -0.788794f, -1.204819f,
74 -1.998864f, -1.182102f, -0.892110f, -1.317415f, -1.359112f, -1.522867f,
75 -1.468552f, -1.779072f, -2.332959f, -2.160346f, -2.329387f, -2.631259f,
76 -2.744936f, -3.052494f, -2.787363f, -3.442548f, -4.245075f, -3.032172f,
77 -2.061609f, -1.768116f, -1.286072f, -0.706587f, -0.192413f, 0.386938f,
78 0.716997f, 1.481393f, 2.216702f, 2.737986f, 3.109809f, 3.226084f,
79 2.490098f, -0.095827f, -3.864816f, -3.507248f, -3.128925f, -2.908251f,
80 -2.883836f, -2.881411f, -2.524377f, -2.624478f, -2.399573f, -2.367718f,
81 -1.918255f, -1.926277f, -1.694584f, -1.723790f, -0.966491f, -1.183115f,
82 -1.430687f, 0.872896f, 2.766550f, 3.610080f, 3.578041f, 3.334928f,
83 2.586680f, 1.895721f, 1.122195f, 0.488519f, -0.140689f, -0.799076f,
84 -1.222860f, -1.502437f, -1.900969f, -3.206816f,
85};
86
87static const NN_CONFIG av1_intra_hog_model_nnconfig = {
88 BINS, // num_inputs
89 DIRECTIONAL_MODES, // num_outputs
90 0, // num_hidden_layers
91 { 0 },
92 {
93 av1_intra_hog_model_weights,
94 },
95 {
96 av1_intra_hog_model_bias,
97 },
98};
99
100#define FIX_PREC_BITS (16)
101static AOM_INLINE int get_hist_bin_idx(int dx, int dy) {
102 const int32_t ratio = (dy * (1 << FIX_PREC_BITS)) / dx;
103
104 // Find index by bisection
105 static const int thresholds[BINS] = {
106 -1334015, -441798, -261605, -183158, -138560, -109331, -88359, -72303,
107 -59392, -48579, -39272, -30982, -23445, -16400, -9715, -3194,
108 3227, 9748, 16433, 23478, 31015, 39305, 48611, 59425,
109 72336, 88392, 109364, 138593, 183191, 261638, 441831, INT32_MAX
110 };
111
112 int lo_idx = 0, hi_idx = BINS - 1;
113 // Divide into segments of size 8 gives better performance than binary search
114 // here.
115 if (ratio <= thresholds[7]) {
116 lo_idx = 0;
117 hi_idx = 7;
118 } else if (ratio <= thresholds[15]) {
119 lo_idx = 8;
120 hi_idx = 15;
121 } else if (ratio <= thresholds[23]) {
122 lo_idx = 16;
123 hi_idx = 23;
124 } else {
125 lo_idx = 24;
126 hi_idx = 31;
127 }
128
129 for (int idx = lo_idx; idx <= hi_idx; idx++) {
130 if (ratio <= thresholds[idx]) {
131 return idx;
132 }
133 }
134 assert(0 && "No valid histogram bin found!");
135 return BINS - 1;
136}
137#undef FIX_PREC_BITS
138
139// Normalizes the hog data.
140static AOM_INLINE void normalize_hog(float total, float *hist) {
141 for (int i = 0; i < BINS; ++i) hist[i] /= total;
142}
143
144static AOM_INLINE void lowbd_generate_hog(const uint8_t *src, int stride,
145 int rows, int cols, float *hist) {
146 float total = 0.1f;
147 src += stride;
148 for (int r = 1; r < rows - 1; ++r) {
149 for (int c = 1; c < cols - 1; ++c) {
150 const uint8_t *above = &src[c - stride];
151 const uint8_t *below = &src[c + stride];
152 const uint8_t *left = &src[c - 1];
153 const uint8_t *right = &src[c + 1];
154 // Calculate gradient using Sobel filters.
155 const int dx = (right[-stride] + 2 * right[0] + right[stride]) -
156 (left[-stride] + 2 * left[0] + left[stride]);
157 const int dy = (below[-1] + 2 * below[0] + below[1]) -
158 (above[-1] + 2 * above[0] + above[1]);
159 if (dx == 0 && dy == 0) continue;
160 const int temp = abs(dx) + abs(dy);
161 if (!temp) continue;
162 total += temp;
163 if (dx == 0) {
164 hist[0] += temp / 2;
165 hist[BINS - 1] += temp / 2;
166 } else {
167 const int idx = get_hist_bin_idx(dx, dy);
168 assert(idx >= 0 && idx < BINS);
169 hist[idx] += temp;
170 }
171 }
172 src += stride;
173 }
174
175 normalize_hog(total, hist);
176}
177
178// Computes and stores pixel level gradient information of a given superblock
179// for LBD encode.
180static AOM_INLINE void lowbd_compute_gradient_info_sb(MACROBLOCK *const x,
181 BLOCK_SIZE sb_size,
182 PLANE_TYPE plane) {
183 PixelLevelGradientInfo *const grad_info_sb =
184 x->pixel_gradient_info + plane * MAX_SB_SQUARE;
185 const uint8_t *src = x->plane[plane].src.buf;
186 const int stride = x->plane[plane].src.stride;
187 const int ss_x = x->e_mbd.plane[plane].subsampling_x;
188 const int ss_y = x->e_mbd.plane[plane].subsampling_y;
189 const int sb_height = block_size_high[sb_size] >> ss_y;
190 const int sb_width = block_size_wide[sb_size] >> ss_x;
191 src += stride;
192 for (int r = 1; r < sb_height - 1; ++r) {
193 for (int c = 1; c < sb_width - 1; ++c) {
194 const uint8_t *above = &src[c - stride];
195 const uint8_t *below = &src[c + stride];
196 const uint8_t *left = &src[c - 1];
197 const uint8_t *right = &src[c + 1];
198 // Calculate gradient using Sobel filters.
199 const int dx = (right[-stride] + 2 * right[0] + right[stride]) -
200 (left[-stride] + 2 * left[0] + left[stride]);
201 const int dy = (below[-1] + 2 * below[0] + below[1]) -
202 (above[-1] + 2 * above[0] + above[1]);
203 grad_info_sb[r * sb_width + c].is_dx_zero = (dx == 0);
204 grad_info_sb[r * sb_width + c].abs_dx_abs_dy_sum =
205 (uint16_t)(abs(dx) + abs(dy));
206 grad_info_sb[r * sb_width + c].hist_bin_idx =
207 (dx != 0) ? get_hist_bin_idx(dx, dy) : -1;
208 }
209 src += stride;
210 }
211}
212
213#if CONFIG_AV1_HIGHBITDEPTH
214static AOM_INLINE void highbd_generate_hog(const uint8_t *src8, int stride,
215 int rows, int cols, float *hist) {
216 float total = 0.1f;
217 const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
218 src += stride;
219 for (int r = 1; r < rows - 1; ++r) {
220 for (int c = 1; c < cols - 1; ++c) {
221 const uint16_t *above = &src[c - stride];
222 const uint16_t *below = &src[c + stride];
223 const uint16_t *left = &src[c - 1];
224 const uint16_t *right = &src[c + 1];
225 // Calculate gradient using Sobel filters.
226 const int dx = (right[-stride] + 2 * right[0] + right[stride]) -
227 (left[-stride] + 2 * left[0] + left[stride]);
228 const int dy = (below[-1] + 2 * below[0] + below[1]) -
229 (above[-1] + 2 * above[0] + above[1]);
230 if (dx == 0 && dy == 0) continue;
231 const int temp = abs(dx) + abs(dy);
232 if (!temp) continue;
233 total += temp;
234 if (dx == 0) {
235 hist[0] += temp / 2;
236 hist[BINS - 1] += temp / 2;
237 } else {
238 const int idx = get_hist_bin_idx(dx, dy);
239 assert(idx >= 0 && idx < BINS);
240 hist[idx] += temp;
241 }
242 }
243 src += stride;
244 }
245
246 normalize_hog(total, hist);
247}
248
249// Computes and stores pixel level gradient information of a given superblock
250// for HBD encode.
251static AOM_INLINE void highbd_compute_gradient_info_sb(MACROBLOCK *const x,
252 BLOCK_SIZE sb_size,
253 PLANE_TYPE plane) {
254 PixelLevelGradientInfo *const grad_info_sb =
255 x->pixel_gradient_info + plane * MAX_SB_SQUARE;
256 const uint16_t *src = CONVERT_TO_SHORTPTR(x->plane[plane].src.buf);
257 const int stride = x->plane[plane].src.stride;
258 const int ss_x = x->e_mbd.plane[plane].subsampling_x;
259 const int ss_y = x->e_mbd.plane[plane].subsampling_y;
260 const int sb_height = block_size_high[sb_size] >> ss_y;
261 const int sb_width = block_size_wide[sb_size] >> ss_x;
262 src += stride;
263 for (int r = 1; r < sb_height - 1; ++r) {
264 for (int c = 1; c < sb_width - 1; ++c) {
265 const uint16_t *above = &src[c - stride];
266 const uint16_t *below = &src[c + stride];
267 const uint16_t *left = &src[c - 1];
268 const uint16_t *right = &src[c + 1];
269 // Calculate gradient using Sobel filters.
270 const int dx = (right[-stride] + 2 * right[0] + right[stride]) -
271 (left[-stride] + 2 * left[0] + left[stride]);
272 const int dy = (below[-1] + 2 * below[0] + below[1]) -
273 (above[-1] + 2 * above[0] + above[1]);
274 grad_info_sb[r * sb_width + c].is_dx_zero = (dx == 0);
275 grad_info_sb[r * sb_width + c].abs_dx_abs_dy_sum =
276 (uint16_t)(abs(dx) + abs(dy));
277 grad_info_sb[r * sb_width + c].hist_bin_idx =
278 (dx != 0) ? get_hist_bin_idx(dx, dy) : -1;
279 }
280 src += stride;
281 }
282}
283#endif // CONFIG_AV1_HIGHBITDEPTH
284
285static AOM_INLINE void generate_hog(const uint8_t *src8, int stride, int rows,
286 int cols, float *hist, int highbd) {
287#if CONFIG_AV1_HIGHBITDEPTH
288 if (highbd) {
289 highbd_generate_hog(src8, stride, rows, cols, hist);
290 return;
291 }
292#else
293 (void)highbd;
294#endif // CONFIG_AV1_HIGHBITDEPTH
295 lowbd_generate_hog(src8, stride, rows, cols, hist);
296}
297
298static AOM_INLINE void compute_gradient_info_sb(MACROBLOCK *const x,
299 BLOCK_SIZE sb_size,
300 PLANE_TYPE plane) {
301#if CONFIG_AV1_HIGHBITDEPTH
302 if (is_cur_buf_hbd(&x->e_mbd)) {
303 highbd_compute_gradient_info_sb(x, sb_size, plane);
304 return;
305 }
306#endif // CONFIG_AV1_HIGHBITDEPTH
307 lowbd_compute_gradient_info_sb(x, sb_size, plane);
308}
309
310// Gradient caching at superblock level is allowed only if all of the following
311// conditions are satisfied:
312// (1) The current frame is an intra only frame
313// (2) Non-RD mode decisions are not enabled
314// (3) The sf partition_search_type is set to SEARCH_PARTITION
315// (4) Either intra_pruning_with_hog or chroma_intra_pruning_with_hog is enabled
316//
317// SB level caching of gradient data may not help in speedup for the following
318// cases:
319// (1) Inter frames (due to early intra gating)
320// (2) When partition_search_type is not SEARCH_PARTITION
321// Hence, gradient data is computed at block level in such cases.
322static AOM_INLINE bool is_gradient_caching_for_hog_enabled(
323 const AV1_COMP *const cpi) {
324 const SPEED_FEATURES *const sf = &cpi->sf;
325 return frame_is_intra_only(&cpi->common) && !sf->rt_sf.use_nonrd_pick_mode &&
326 (sf->part_sf.partition_search_type == SEARCH_PARTITION) &&
327 (sf->intra_sf.intra_pruning_with_hog ||
328 sf->intra_sf.chroma_intra_pruning_with_hog);
329}
330
331// Function to generate pixel level gradient information for a given superblock.
332// Sets the flags 'is_sb_gradient_cached' for the specific plane-type if
333// gradient info is generated for the same.
334static AOM_INLINE void produce_gradients_for_sb(AV1_COMP *cpi, MACROBLOCK *x,
335 BLOCK_SIZE sb_size, int mi_row,
336 int mi_col) {
337 // Initialise flags related to hog data caching.
338 x->is_sb_gradient_cached[PLANE_TYPE_Y] = false;
339 x->is_sb_gradient_cached[PLANE_TYPE_UV] = false;
340 if (!is_gradient_caching_for_hog_enabled(cpi)) return;
341
342 const SPEED_FEATURES *sf = &cpi->sf;
343 const int num_planes = av1_num_planes(&cpi->common);
344
345 av1_setup_src_planes(x, cpi->source, mi_row, mi_col, num_planes, sb_size);
346
347 if (sf->intra_sf.intra_pruning_with_hog) {
348 compute_gradient_info_sb(x, sb_size, PLANE_TYPE_Y);
349 x->is_sb_gradient_cached[PLANE_TYPE_Y] = true;
350 }
351 if (sf->intra_sf.chroma_intra_pruning_with_hog && num_planes > 1) {
352 compute_gradient_info_sb(x, sb_size, PLANE_TYPE_UV);
353 x->is_sb_gradient_cached[PLANE_TYPE_UV] = true;
354 }
355}
356
357// Reuses the pixel level gradient data generated at superblock level for block
358// level histogram computation.
359static AOM_INLINE void generate_hog_using_gradient_cache(const MACROBLOCK *x,
360 int rows, int cols,
361 BLOCK_SIZE sb_size,
362 PLANE_TYPE plane,
363 float *hist) {
364 float total = 0.1f;
365 const int ss_x = x->e_mbd.plane[plane].subsampling_x;
366 const int ss_y = x->e_mbd.plane[plane].subsampling_y;
367 const int sb_width = block_size_wide[sb_size] >> ss_x;
368
369 // Derive the offset from the starting of the superblock in order to locate
370 // the block level gradient data in the cache.
371 const int mi_row_in_sb = x->e_mbd.mi_row & (mi_size_high[sb_size] - 1);
372 const int mi_col_in_sb = x->e_mbd.mi_col & (mi_size_wide[sb_size] - 1);
373 const int block_offset_in_grad_cache =
374 sb_width * (mi_row_in_sb << (MI_SIZE_LOG2 - ss_y)) +
375 (mi_col_in_sb << (MI_SIZE_LOG2 - ss_x));
376 const PixelLevelGradientInfo *grad_info_blk = x->pixel_gradient_info +
377 plane * MAX_SB_SQUARE +
378 block_offset_in_grad_cache;
379
380 // Retrieve the cached gradient information and generate the histogram.
381 for (int r = 1; r < rows - 1; ++r) {
382 for (int c = 1; c < cols - 1; ++c) {
383 const uint16_t abs_dx_abs_dy_sum =
384 grad_info_blk[r * sb_width + c].abs_dx_abs_dy_sum;
385 if (!abs_dx_abs_dy_sum) continue;
386 total += abs_dx_abs_dy_sum;
387 const bool is_dx_zero = grad_info_blk[r * sb_width + c].is_dx_zero;
388 if (is_dx_zero) {
389 hist[0] += abs_dx_abs_dy_sum >> 1;
390 hist[BINS - 1] += abs_dx_abs_dy_sum >> 1;
391 } else {
392 const int8_t idx = grad_info_blk[r * sb_width + c].hist_bin_idx;
393 assert(idx >= 0 && idx < BINS);
394 hist[idx] += abs_dx_abs_dy_sum;
395 }
396 }
397 }
398 normalize_hog(total, hist);
399}
400
401static INLINE void collect_hog_data(const MACROBLOCK *x, BLOCK_SIZE bsize,
402 BLOCK_SIZE sb_size, int plane, float *hog) {
403 const MACROBLOCKD *xd = &x->e_mbd;
404 const struct macroblockd_plane *const pd = &xd->plane[plane];
405 const int ss_x = pd->subsampling_x;
406 const int ss_y = pd->subsampling_y;
407 const int bh = block_size_high[bsize];
408 const int bw = block_size_wide[bsize];
409 const int rows =
410 ((xd->mb_to_bottom_edge >= 0) ? bh : (xd->mb_to_bottom_edge >> 3) + bh) >>
411 ss_y;
412 const int cols =
413 ((xd->mb_to_right_edge >= 0) ? bw : (xd->mb_to_right_edge >> 3) + bw) >>
414 ss_x;
415
416 // If gradient data is already generated at SB level, reuse the cached data.
417 // Otherwise, compute the data.
418 if (x->is_sb_gradient_cached[plane]) {
419 generate_hog_using_gradient_cache(x, rows, cols, sb_size, plane, hog);
420 } else {
421 const uint8_t *src = x->plane[plane].src.buf;
422 const int src_stride = x->plane[plane].src.stride;
423 generate_hog(src, src_stride, rows, cols, hog, is_cur_buf_hbd(xd));
424 }
425
426 // Scale the hog so the luma and chroma are on the same scale
427 for (int b = 0; b < BINS; ++b) {
428 hog[b] *= (1 + ss_x) * (1 + ss_y);
429 }
430}
431
432static AOM_INLINE void prune_intra_mode_with_hog(
433 const MACROBLOCK *x, BLOCK_SIZE bsize, BLOCK_SIZE sb_size, float th,
434 uint8_t *directional_mode_skip_mask, int is_chroma) {
435 const int plane = is_chroma ? AOM_PLANE_U : AOM_PLANE_Y;
436 float hist[BINS] = { 0.0f };
437 collect_hog_data(x, bsize, sb_size, plane, hist);
438
439 // Make prediction for each of the mode
440 float scores[DIRECTIONAL_MODES] = { 0.0f };
441 av1_nn_predict(hist, &av1_intra_hog_model_nnconfig, 1, scores);
442 for (UV_PREDICTION_MODE uv_mode = UV_V_PRED; uv_mode <= UV_D67_PRED;
443 uv_mode++) {
444 if (scores[uv_mode - UV_V_PRED] <= th) {
445 directional_mode_skip_mask[uv_mode] = 1;
446 }
447 }
448}
449#undef BINS
450
451// Returns the cost needed to send a uniformly distributed r.v.
452static AOM_INLINE int write_uniform_cost(int n, int v) {
453 const int l = get_unsigned_bits(n);
454 const int m = (1 << l) - n;
455 if (l == 0) return 0;
456 if (v < m)
457 return av1_cost_literal(l - 1);
458 else
459 return av1_cost_literal(l);
460}
467static AOM_INLINE int intra_mode_info_cost_y(const AV1_COMP *cpi,
468 const MACROBLOCK *x,
469 const MB_MODE_INFO *mbmi,
470 BLOCK_SIZE bsize, int mode_cost) {
471 int total_rate = mode_cost;
472 const ModeCosts *mode_costs = &x->mode_costs;
473 const int use_palette = mbmi->palette_mode_info.palette_size[0] > 0;
474 const int use_filter_intra = mbmi->filter_intra_mode_info.use_filter_intra;
475 const int use_intrabc = mbmi->use_intrabc;
476 // Can only activate one mode.
477 assert(((mbmi->mode != DC_PRED) + use_palette + use_intrabc +
478 use_filter_intra) <= 1);
479 const int try_palette = av1_allow_palette(
480 cpi->common.features.allow_screen_content_tools, mbmi->bsize);
481 if (try_palette && mbmi->mode == DC_PRED) {
482 const MACROBLOCKD *xd = &x->e_mbd;
483 const int bsize_ctx = av1_get_palette_bsize_ctx(bsize);
484 const int mode_ctx = av1_get_palette_mode_ctx(xd);
485 total_rate +=
486 mode_costs->palette_y_mode_cost[bsize_ctx][mode_ctx][use_palette];
487 if (use_palette) {
488 const uint8_t *const color_map = xd->plane[0].color_index_map;
489 int block_width, block_height, rows, cols;
490 av1_get_block_dimensions(bsize, 0, xd, &block_width, &block_height, &rows,
491 &cols);
492 const int plt_size = mbmi->palette_mode_info.palette_size[0];
493 int palette_mode_cost =
494 mode_costs
495 ->palette_y_size_cost[bsize_ctx][plt_size - PALETTE_MIN_SIZE] +
496 write_uniform_cost(plt_size, color_map[0]);
497 uint16_t color_cache[2 * PALETTE_MAX_SIZE];
498 const int n_cache = av1_get_palette_cache(xd, 0, color_cache);
499 palette_mode_cost +=
501 n_cache, cpi->common.seq_params->bit_depth);
502 palette_mode_cost +=
503 av1_cost_color_map(x, 0, bsize, mbmi->tx_size, PALETTE_MAP);
504 total_rate += palette_mode_cost;
505 }
506 }
507 if (av1_filter_intra_allowed(&cpi->common, mbmi)) {
508 total_rate += mode_costs->filter_intra_cost[mbmi->bsize][use_filter_intra];
509 if (use_filter_intra) {
510 total_rate +=
512 .filter_intra_mode];
513 }
514 }
515 if (av1_is_directional_mode(mbmi->mode)) {
516 if (av1_use_angle_delta(bsize)) {
517 total_rate +=
518 mode_costs->angle_delta_cost[mbmi->mode - V_PRED]
519 [MAX_ANGLE_DELTA +
520 mbmi->angle_delta[PLANE_TYPE_Y]];
521 }
522 }
523 if (av1_allow_intrabc(&cpi->common))
524 total_rate += mode_costs->intrabc_cost[use_intrabc];
525 return total_rate;
526}
527
532static AOM_INLINE int intra_mode_info_cost_uv(const AV1_COMP *cpi,
533 const MACROBLOCK *x,
534 const MB_MODE_INFO *mbmi,
535 BLOCK_SIZE bsize, int mode_cost) {
536 int total_rate = mode_cost;
537 const ModeCosts *mode_costs = &x->mode_costs;
538 const int use_palette = mbmi->palette_mode_info.palette_size[1] > 0;
539 const UV_PREDICTION_MODE mode = mbmi->uv_mode;
540 // Can only activate one mode.
541 assert(((mode != UV_DC_PRED) + use_palette + mbmi->use_intrabc) <= 1);
542
543 const int try_palette = av1_allow_palette(
544 cpi->common.features.allow_screen_content_tools, mbmi->bsize);
545 if (try_palette && mode == UV_DC_PRED) {
546 const PALETTE_MODE_INFO *pmi = &mbmi->palette_mode_info;
547 total_rate +=
548 mode_costs->palette_uv_mode_cost[pmi->palette_size[0] > 0][use_palette];
549 if (use_palette) {
550 const int bsize_ctx = av1_get_palette_bsize_ctx(bsize);
551 const int plt_size = pmi->palette_size[1];
552 const MACROBLOCKD *xd = &x->e_mbd;
553 const uint8_t *const color_map = xd->plane[1].color_index_map;
554 int palette_mode_cost =
555 mode_costs
556 ->palette_uv_size_cost[bsize_ctx][plt_size - PALETTE_MIN_SIZE] +
557 write_uniform_cost(plt_size, color_map[0]);
558 uint16_t color_cache[2 * PALETTE_MAX_SIZE];
559 const int n_cache = av1_get_palette_cache(xd, 1, color_cache);
560 palette_mode_cost += av1_palette_color_cost_uv(
561 pmi, color_cache, n_cache, cpi->common.seq_params->bit_depth);
562 palette_mode_cost +=
563 av1_cost_color_map(x, 1, bsize, mbmi->tx_size, PALETTE_MAP);
564 total_rate += palette_mode_cost;
565 }
566 }
567 if (av1_is_directional_mode(get_uv_mode(mode))) {
568 if (av1_use_angle_delta(bsize)) {
569 total_rate +=
570 mode_costs->angle_delta_cost[mode - V_PRED]
571 [mbmi->angle_delta[PLANE_TYPE_UV] +
572 MAX_ANGLE_DELTA];
573 }
574 }
575 return total_rate;
576}
577
579// Makes a quick intra prediction and estimate the rdcost with a model without
580// going through the whole txfm/quantize/itxfm process.
581static int64_t intra_model_rd(const AV1_COMMON *cm, MACROBLOCK *const x,
582 int plane, BLOCK_SIZE plane_bsize,
583 TX_SIZE tx_size, int use_hadamard) {
584 MACROBLOCKD *const xd = &x->e_mbd;
585 const BitDepthInfo bd_info = get_bit_depth_info(xd);
586 int row, col;
587 assert(!is_inter_block(xd->mi[0]));
588 const int stepr = tx_size_high_unit[tx_size];
589 const int stepc = tx_size_wide_unit[tx_size];
590 const int txbw = tx_size_wide[tx_size];
591 const int txbh = tx_size_high[tx_size];
592 const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
593 const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
594 int64_t satd_cost = 0;
595 struct macroblock_plane *p = &x->plane[plane];
596 struct macroblockd_plane *pd = &xd->plane[plane];
597 // Prediction.
598 for (row = 0; row < max_blocks_high; row += stepr) {
599 for (col = 0; col < max_blocks_wide; col += stepc) {
600 av1_predict_intra_block_facade(cm, xd, plane, col, row, tx_size);
601 // Here we use p->src_diff and p->coeff as temporary buffers for
602 // prediction residue and transform coefficients. The buffers are only
603 // used in this for loop, therefore we don't need to properly add offset
604 // to the buffers.
605 av1_subtract_block(
606 bd_info, txbh, txbw, p->src_diff, block_size_wide[plane_bsize],
607 p->src.buf + (((row * p->src.stride) + col) << 2), p->src.stride,
608 pd->dst.buf + (((row * pd->dst.stride) + col) << 2), pd->dst.stride);
609 av1_quick_txfm(use_hadamard, tx_size, bd_info, p->src_diff,
610 block_size_wide[plane_bsize], p->coeff);
611 satd_cost += aom_satd(p->coeff, tx_size_2d[tx_size]);
612 }
613 }
614 return satd_cost;
615}
628static AOM_INLINE int model_intra_yrd_and_prune(const AV1_COMP *const cpi,
629 MACROBLOCK *x, BLOCK_SIZE bsize,
630 int64_t *best_model_rd) {
631 const TX_SIZE tx_size = AOMMIN(TX_32X32, max_txsize_lookup[bsize]);
632 const int plane = 0;
633 const AV1_COMMON *cm = &cpi->common;
634 const int64_t this_model_rd =
635 intra_model_rd(cm, x, plane, bsize, tx_size, /*use_hadamard=*/1);
636 if (*best_model_rd != INT64_MAX &&
637 this_model_rd > *best_model_rd + (*best_model_rd >> 2)) {
638 return 1;
639 } else if (this_model_rd < *best_model_rd) {
640 *best_model_rd = this_model_rd;
641 }
642 return 0;
643}
644
645#ifdef __cplusplus
646} // extern "C"
647#endif
648
649#endif // AOM_AV1_ENCODER_INTRA_MODE_SEARCH_UTILS_H_
#define AOM_PLANE_U
Definition: aom_image.h:200
#define AOM_PLANE_Y
Definition: aom_image.h:199
Declares top-level encoder structures and functions.
static int model_intra_yrd_and_prune(const AV1_COMP *const cpi, MACROBLOCK *x, BLOCK_SIZE bsize, int64_t *best_model_rd)
Estimate the luma rdcost of a given intra mode and try to prune it.
Definition: intra_mode_search_utils.h:628
int av1_palette_color_cost_y(const PALETTE_MODE_INFO *const pmi, const uint16_t *color_cache, int n_cache, int bit_depth)
Gets the rate cost for transmitting luma palette color values.
Definition: palette.c:125
int av1_palette_color_cost_uv(const PALETTE_MODE_INFO *const pmi, const uint16_t *color_cache, int n_cache, int bit_depth)
Gets the rate cost for transmitting luma palette chroma values.
Definition: palette.c:139
static int intra_mode_info_cost_uv(const AV1_COMP *cpi, const MACROBLOCK *x, const MB_MODE_INFO *mbmi, BLOCK_SIZE bsize, int mode_cost)
Return the rate cost for chroma prediction mode info of intra blocks.
Definition: intra_mode_search_utils.h:532
static int intra_mode_info_cost_y(const AV1_COMP *cpi, const MACROBLOCK *x, const MB_MODE_INFO *mbmi, BLOCK_SIZE bsize, int mode_cost)
Returns the rate cost for luma prediction mode info of intra blocks.
Definition: intra_mode_search_utils.h:467
Declares functions used in palette search.
Top level common structure used by both encoder and decoder.
Definition: av1_common_int.h:755
Stores the prediction/txfm mode of the current coding block.
Definition: blockd.h:222
PREDICTION_MODE mode
The prediction mode used.
Definition: blockd.h:232
UV_PREDICTION_MODE uv_mode
The UV mode when intra is used.
Definition: blockd.h:234
PALETTE_MODE_INFO palette_mode_info
Stores the size and colors of palette mode.
Definition: blockd.h:280
int8_t angle_delta[PLANE_TYPES]
Directional mode delta: the angle is base angle + (angle_delta * step).
Definition: blockd.h:272
FILTER_INTRA_MODE_INFO filter_intra_mode_info
The type of filter intra mode used (if applicable).
Definition: blockd.h:274
BLOCK_SIZE bsize
The block size of the current coding block.
Definition: blockd.h:228
TX_SIZE tx_size
Transform size when fixed size txfm is used (e.g. intra modes).
Definition: blockd.h:290
uint8_t use_intrabc
Whether intrabc is used.
Definition: blockd.h:318
Holds the entropy costs for various modes sent to the bitstream.
Definition: block.h:507
int intrabc_cost[2]
intrabc_cost
Definition: block.h:542
int palette_uv_mode_cost[2][2]
palette_uv_mode_cost
Definition: block.h:557
int palette_y_size_cost[7][PALETTE_SIZES]
palette_y_size_cost
Definition: block.h:545
int palette_uv_size_cost[7][PALETTE_SIZES]
palette_uv_size_cost
Definition: block.h:547
int filter_intra_cost[BLOCK_SIZES_ALL][2]
filter_intra_cost
Definition: block.h:527
int filter_intra_mode_cost[FILTER_INTRA_MODES]
filter_intra_mode_cost
Definition: block.h:529
int angle_delta_cost[DIRECTIONAL_MODES][2 *MAX_ANGLE_DELTA+1]
angle_delta_cost
Definition: block.h:531
int palette_y_mode_cost[7][3][2]
palette_y_mode_cost
Definition: block.h:555
Top level speed vs quality trade off data struture.
Definition: speed_features.h:1388
PARTITION_SPEED_FEATURES part_sf
Definition: speed_features.h:1412
INTRA_MODE_SPEED_FEATURES intra_sf
Definition: speed_features.h:1432
REAL_TIME_SPEED_FEATURES rt_sf
Definition: speed_features.h:1457
Each source plane of the current macroblock.
Definition: block.h:103
struct buf_2d src
A buffer containing the source frame.
Definition: block.h:117
tran_low_t * coeff
Transformed coefficients.
Definition: block.h:111
int16_t * src_diff
Stores source - pred so the txfm can be computed later.
Definition: block.h:105
Encoder's parameters related to the current coding block.
Definition: block.h:778
MACROBLOCKD e_mbd
Decoder's view of current coding block.
Definition: block.h:796
ModeCosts mode_costs
The rate needed to signal a mode to the bitstream.
Definition: block.h:891
PixelLevelGradientInfo * pixel_gradient_info
Pointer to the buffer which caches gradient information.
Definition: block.h:1118
struct macroblock_plane plane[3]
Each of the encoding plane.
Definition: block.h:788
bool is_sb_gradient_cached[PLANE_TYPES]
Flags indicating the availability of cached gradient info.
Definition: block.h:1120
Variables related to current coding block.
Definition: blockd.h:577
int mb_to_bottom_edge
Definition: blockd.h:687
struct macroblockd_plane plane[3]
Definition: blockd.h:613
int mb_to_right_edge
Definition: blockd.h:685
int mi_row
Definition: blockd.h:582
MB_MODE_INFO ** mi
Definition: blockd.h:624
int mi_col
Definition: blockd.h:583