1 /*
2 * Copyright (c) 2016, 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 #include "av1/common/cfl.h"
13 #include "av1/common/common_data.h"
14 #include "av1/common/onyxc_int.h"
15
16 #include "config/av1_rtcd.h"
17
cfl_init(CFL_CTX * cfl,const SequenceHeader * seq_params)18 void cfl_init(CFL_CTX *cfl, const SequenceHeader *seq_params) {
19 assert(block_size_wide[CFL_MAX_BLOCK_SIZE] == CFL_BUF_LINE);
20 assert(block_size_high[CFL_MAX_BLOCK_SIZE] == CFL_BUF_LINE);
21
22 memset(&cfl->recon_buf_q3, 0, sizeof(cfl->recon_buf_q3));
23 memset(&cfl->ac_buf_q3, 0, sizeof(cfl->ac_buf_q3));
24 cfl->subsampling_x = seq_params->subsampling_x;
25 cfl->subsampling_y = seq_params->subsampling_y;
26 cfl->are_parameters_computed = 0;
27 cfl->store_y = 0;
28 // The DC_PRED cache is disabled by default and is only enabled in
29 // cfl_rd_pick_alpha
30 cfl->use_dc_pred_cache = 0;
31 cfl->dc_pred_is_cached[CFL_PRED_U] = 0;
32 cfl->dc_pred_is_cached[CFL_PRED_V] = 0;
33 }
34
cfl_store_dc_pred(MACROBLOCKD * const xd,const uint8_t * input,CFL_PRED_TYPE pred_plane,int width)35 void cfl_store_dc_pred(MACROBLOCKD *const xd, const uint8_t *input,
36 CFL_PRED_TYPE pred_plane, int width) {
37 assert(pred_plane < CFL_PRED_PLANES);
38 assert(width <= CFL_BUF_LINE);
39
40 if (get_bitdepth_data_path_index(xd)) {
41 uint16_t *const input_16 = CONVERT_TO_SHORTPTR(input);
42 memcpy(xd->cfl.dc_pred_cache[pred_plane], input_16, width << 1);
43 return;
44 }
45
46 memcpy(xd->cfl.dc_pred_cache[pred_plane], input, width);
47 }
48
cfl_load_dc_pred_lbd(const int16_t * dc_pred_cache,uint8_t * dst,int dst_stride,int width,int height)49 static void cfl_load_dc_pred_lbd(const int16_t *dc_pred_cache, uint8_t *dst,
50 int dst_stride, int width, int height) {
51 for (int j = 0; j < height; j++) {
52 memcpy(dst, dc_pred_cache, width);
53 dst += dst_stride;
54 }
55 }
56
cfl_load_dc_pred_hbd(const int16_t * dc_pred_cache,uint16_t * dst,int dst_stride,int width,int height)57 static void cfl_load_dc_pred_hbd(const int16_t *dc_pred_cache, uint16_t *dst,
58 int dst_stride, int width, int height) {
59 const size_t num_bytes = width << 1;
60 for (int j = 0; j < height; j++) {
61 memcpy(dst, dc_pred_cache, num_bytes);
62 dst += dst_stride;
63 }
64 }
cfl_load_dc_pred(MACROBLOCKD * const xd,uint8_t * dst,int dst_stride,TX_SIZE tx_size,CFL_PRED_TYPE pred_plane)65 void cfl_load_dc_pred(MACROBLOCKD *const xd, uint8_t *dst, int dst_stride,
66 TX_SIZE tx_size, CFL_PRED_TYPE pred_plane) {
67 const int width = tx_size_wide[tx_size];
68 const int height = tx_size_high[tx_size];
69 assert(pred_plane < CFL_PRED_PLANES);
70 assert(width <= CFL_BUF_LINE);
71 assert(height <= CFL_BUF_LINE);
72 if (get_bitdepth_data_path_index(xd)) {
73 uint16_t *dst_16 = CONVERT_TO_SHORTPTR(dst);
74 cfl_load_dc_pred_hbd(xd->cfl.dc_pred_cache[pred_plane], dst_16, dst_stride,
75 width, height);
76 return;
77 }
78 cfl_load_dc_pred_lbd(xd->cfl.dc_pred_cache[pred_plane], dst, dst_stride,
79 width, height);
80 }
81
82 // Due to frame boundary issues, it is possible that the total area covered by
83 // chroma exceeds that of luma. When this happens, we fill the missing pixels by
84 // repeating the last columns and/or rows.
cfl_pad(CFL_CTX * cfl,int width,int height)85 static INLINE void cfl_pad(CFL_CTX *cfl, int width, int height) {
86 const int diff_width = width - cfl->buf_width;
87 const int diff_height = height - cfl->buf_height;
88
89 if (diff_width > 0) {
90 const int min_height = height - diff_height;
91 uint16_t *recon_buf_q3 = cfl->recon_buf_q3 + (width - diff_width);
92 for (int j = 0; j < min_height; j++) {
93 const uint16_t last_pixel = recon_buf_q3[-1];
94 assert(recon_buf_q3 + diff_width <= cfl->recon_buf_q3 + CFL_BUF_SQUARE);
95 for (int i = 0; i < diff_width; i++) {
96 recon_buf_q3[i] = last_pixel;
97 }
98 recon_buf_q3 += CFL_BUF_LINE;
99 }
100 cfl->buf_width = width;
101 }
102 if (diff_height > 0) {
103 uint16_t *recon_buf_q3 =
104 cfl->recon_buf_q3 + ((height - diff_height) * CFL_BUF_LINE);
105 for (int j = 0; j < diff_height; j++) {
106 const uint16_t *last_row_q3 = recon_buf_q3 - CFL_BUF_LINE;
107 assert(recon_buf_q3 + width <= cfl->recon_buf_q3 + CFL_BUF_SQUARE);
108 for (int i = 0; i < width; i++) {
109 recon_buf_q3[i] = last_row_q3[i];
110 }
111 recon_buf_q3 += CFL_BUF_LINE;
112 }
113 cfl->buf_height = height;
114 }
115 }
116
subtract_average_c(const uint16_t * src,int16_t * dst,int width,int height,int round_offset,int num_pel_log2)117 static void subtract_average_c(const uint16_t *src, int16_t *dst, int width,
118 int height, int round_offset, int num_pel_log2) {
119 int sum = round_offset;
120 const uint16_t *recon = src;
121 for (int j = 0; j < height; j++) {
122 for (int i = 0; i < width; i++) {
123 sum += recon[i];
124 }
125 recon += CFL_BUF_LINE;
126 }
127 const int avg = sum >> num_pel_log2;
128 for (int j = 0; j < height; j++) {
129 for (int i = 0; i < width; i++) {
130 dst[i] = src[i] - avg;
131 }
132 src += CFL_BUF_LINE;
133 dst += CFL_BUF_LINE;
134 }
135 }
136
CFL_SUB_AVG_FN(c)137 CFL_SUB_AVG_FN(c)
138
139 static INLINE int cfl_idx_to_alpha(int alpha_idx, int joint_sign,
140 CFL_PRED_TYPE pred_type) {
141 const int alpha_sign = (pred_type == CFL_PRED_U) ? CFL_SIGN_U(joint_sign)
142 : CFL_SIGN_V(joint_sign);
143 if (alpha_sign == CFL_SIGN_ZERO) return 0;
144 const int abs_alpha_q3 =
145 (pred_type == CFL_PRED_U) ? CFL_IDX_U(alpha_idx) : CFL_IDX_V(alpha_idx);
146 return (alpha_sign == CFL_SIGN_POS) ? abs_alpha_q3 + 1 : -abs_alpha_q3 - 1;
147 }
148
cfl_predict_lbd_c(const int16_t * ac_buf_q3,uint8_t * dst,int dst_stride,int alpha_q3,int width,int height)149 static INLINE void cfl_predict_lbd_c(const int16_t *ac_buf_q3, uint8_t *dst,
150 int dst_stride, int alpha_q3, int width,
151 int height) {
152 for (int j = 0; j < height; j++) {
153 for (int i = 0; i < width; i++) {
154 dst[i] = clip_pixel(get_scaled_luma_q0(alpha_q3, ac_buf_q3[i]) + dst[i]);
155 }
156 dst += dst_stride;
157 ac_buf_q3 += CFL_BUF_LINE;
158 }
159 }
160
161 // Null function used for invalid tx_sizes
cfl_predict_lbd_null(const int16_t * ac_buf_q3,uint8_t * dst,int dst_stride,int alpha_q3)162 void cfl_predict_lbd_null(const int16_t *ac_buf_q3, uint8_t *dst,
163 int dst_stride, int alpha_q3) {
164 (void)ac_buf_q3;
165 (void)dst;
166 (void)dst_stride;
167 (void)alpha_q3;
168 assert(0);
169 }
170
CFL_PREDICT_FN(c,lbd)171 CFL_PREDICT_FN(c, lbd)
172
173 void cfl_predict_hbd_c(const int16_t *ac_buf_q3, uint16_t *dst, int dst_stride,
174 int alpha_q3, int bit_depth, int width, int height) {
175 for (int j = 0; j < height; j++) {
176 for (int i = 0; i < width; i++) {
177 dst[i] = clip_pixel_highbd(
178 get_scaled_luma_q0(alpha_q3, ac_buf_q3[i]) + dst[i], bit_depth);
179 }
180 dst += dst_stride;
181 ac_buf_q3 += CFL_BUF_LINE;
182 }
183 }
184
185 // Null function used for invalid tx_sizes
cfl_predict_hbd_null(const int16_t * ac_buf_q3,uint16_t * dst,int dst_stride,int alpha_q3,int bd)186 void cfl_predict_hbd_null(const int16_t *ac_buf_q3, uint16_t *dst,
187 int dst_stride, int alpha_q3, int bd) {
188 (void)ac_buf_q3;
189 (void)dst;
190 (void)dst_stride;
191 (void)alpha_q3;
192 (void)bd;
193 assert(0);
194 }
195
CFL_PREDICT_FN(c,hbd)196 CFL_PREDICT_FN(c, hbd)
197
198 static void cfl_compute_parameters(MACROBLOCKD *const xd, TX_SIZE tx_size) {
199 CFL_CTX *const cfl = &xd->cfl;
200 // Do not call cfl_compute_parameters multiple time on the same values.
201 assert(cfl->are_parameters_computed == 0);
202
203 cfl_pad(cfl, tx_size_wide[tx_size], tx_size_high[tx_size]);
204 get_subtract_average_fn(tx_size)(cfl->recon_buf_q3, cfl->ac_buf_q3);
205 cfl->are_parameters_computed = 1;
206 }
207
cfl_predict_block(MACROBLOCKD * const xd,uint8_t * dst,int dst_stride,TX_SIZE tx_size,int plane)208 void cfl_predict_block(MACROBLOCKD *const xd, uint8_t *dst, int dst_stride,
209 TX_SIZE tx_size, int plane) {
210 CFL_CTX *const cfl = &xd->cfl;
211 MB_MODE_INFO *mbmi = xd->mi[0];
212 assert(is_cfl_allowed(xd));
213
214 if (!cfl->are_parameters_computed) cfl_compute_parameters(xd, tx_size);
215
216 const int alpha_q3 =
217 cfl_idx_to_alpha(mbmi->cfl_alpha_idx, mbmi->cfl_alpha_signs, plane - 1);
218 assert((tx_size_high[tx_size] - 1) * CFL_BUF_LINE + tx_size_wide[tx_size] <=
219 CFL_BUF_SQUARE);
220 if (get_bitdepth_data_path_index(xd)) {
221 uint16_t *dst_16 = CONVERT_TO_SHORTPTR(dst);
222 get_predict_hbd_fn(tx_size)(cfl->ac_buf_q3, dst_16, dst_stride, alpha_q3,
223 xd->bd);
224 return;
225 }
226 get_predict_lbd_fn(tx_size)(cfl->ac_buf_q3, dst, dst_stride, alpha_q3);
227 }
228
229 // Null function used for invalid tx_sizes
cfl_subsample_lbd_null(const uint8_t * input,int input_stride,uint16_t * output_q3)230 void cfl_subsample_lbd_null(const uint8_t *input, int input_stride,
231 uint16_t *output_q3) {
232 (void)input;
233 (void)input_stride;
234 (void)output_q3;
235 assert(0);
236 }
237
238 // Null function used for invalid tx_sizes
cfl_subsample_hbd_null(const uint16_t * input,int input_stride,uint16_t * output_q3)239 void cfl_subsample_hbd_null(const uint16_t *input, int input_stride,
240 uint16_t *output_q3) {
241 (void)input;
242 (void)input_stride;
243 (void)output_q3;
244 assert(0);
245 }
246
cfl_luma_subsampling_420_lbd_c(const uint8_t * input,int input_stride,uint16_t * output_q3,int width,int height)247 static void cfl_luma_subsampling_420_lbd_c(const uint8_t *input,
248 int input_stride,
249 uint16_t *output_q3, int width,
250 int height) {
251 for (int j = 0; j < height; j += 2) {
252 for (int i = 0; i < width; i += 2) {
253 const int bot = i + input_stride;
254 output_q3[i >> 1] =
255 (input[i] + input[i + 1] + input[bot] + input[bot + 1]) << 1;
256 }
257 input += input_stride << 1;
258 output_q3 += CFL_BUF_LINE;
259 }
260 }
261
cfl_luma_subsampling_422_lbd_c(const uint8_t * input,int input_stride,uint16_t * output_q3,int width,int height)262 static void cfl_luma_subsampling_422_lbd_c(const uint8_t *input,
263 int input_stride,
264 uint16_t *output_q3, int width,
265 int height) {
266 assert((height - 1) * CFL_BUF_LINE + width <= CFL_BUF_SQUARE);
267 for (int j = 0; j < height; j++) {
268 for (int i = 0; i < width; i += 2) {
269 output_q3[i >> 1] = (input[i] + input[i + 1]) << 2;
270 }
271 input += input_stride;
272 output_q3 += CFL_BUF_LINE;
273 }
274 }
275
cfl_luma_subsampling_444_lbd_c(const uint8_t * input,int input_stride,uint16_t * output_q3,int width,int height)276 static void cfl_luma_subsampling_444_lbd_c(const uint8_t *input,
277 int input_stride,
278 uint16_t *output_q3, int width,
279 int height) {
280 assert((height - 1) * CFL_BUF_LINE + width <= CFL_BUF_SQUARE);
281 for (int j = 0; j < height; j++) {
282 for (int i = 0; i < width; i++) {
283 output_q3[i] = input[i] << 3;
284 }
285 input += input_stride;
286 output_q3 += CFL_BUF_LINE;
287 }
288 }
289
cfl_luma_subsampling_420_hbd_c(const uint16_t * input,int input_stride,uint16_t * output_q3,int width,int height)290 static void cfl_luma_subsampling_420_hbd_c(const uint16_t *input,
291 int input_stride,
292 uint16_t *output_q3, int width,
293 int height) {
294 for (int j = 0; j < height; j += 2) {
295 for (int i = 0; i < width; i += 2) {
296 const int bot = i + input_stride;
297 output_q3[i >> 1] =
298 (input[i] + input[i + 1] + input[bot] + input[bot + 1]) << 1;
299 }
300 input += input_stride << 1;
301 output_q3 += CFL_BUF_LINE;
302 }
303 }
304
cfl_luma_subsampling_422_hbd_c(const uint16_t * input,int input_stride,uint16_t * output_q3,int width,int height)305 static void cfl_luma_subsampling_422_hbd_c(const uint16_t *input,
306 int input_stride,
307 uint16_t *output_q3, int width,
308 int height) {
309 assert((height - 1) * CFL_BUF_LINE + width <= CFL_BUF_SQUARE);
310 for (int j = 0; j < height; j++) {
311 for (int i = 0; i < width; i += 2) {
312 output_q3[i >> 1] = (input[i] + input[i + 1]) << 2;
313 }
314 input += input_stride;
315 output_q3 += CFL_BUF_LINE;
316 }
317 }
318
cfl_luma_subsampling_444_hbd_c(const uint16_t * input,int input_stride,uint16_t * output_q3,int width,int height)319 static void cfl_luma_subsampling_444_hbd_c(const uint16_t *input,
320 int input_stride,
321 uint16_t *output_q3, int width,
322 int height) {
323 assert((height - 1) * CFL_BUF_LINE + width <= CFL_BUF_SQUARE);
324 for (int j = 0; j < height; j++) {
325 for (int i = 0; i < width; i++) {
326 output_q3[i] = input[i] << 3;
327 }
328 input += input_stride;
329 output_q3 += CFL_BUF_LINE;
330 }
331 }
332
CFL_GET_SUBSAMPLE_FUNCTION(c)333 CFL_GET_SUBSAMPLE_FUNCTION(c)
334
335 static INLINE cfl_subsample_hbd_fn cfl_subsampling_hbd(TX_SIZE tx_size,
336 int sub_x, int sub_y) {
337 if (sub_x == 1) {
338 if (sub_y == 1) {
339 return cfl_get_luma_subsampling_420_hbd(tx_size);
340 }
341 return cfl_get_luma_subsampling_422_hbd(tx_size);
342 }
343 return cfl_get_luma_subsampling_444_hbd(tx_size);
344 }
345
cfl_subsampling_lbd(TX_SIZE tx_size,int sub_x,int sub_y)346 static INLINE cfl_subsample_lbd_fn cfl_subsampling_lbd(TX_SIZE tx_size,
347 int sub_x, int sub_y) {
348 if (sub_x == 1) {
349 if (sub_y == 1) {
350 return cfl_get_luma_subsampling_420_lbd(tx_size);
351 }
352 return cfl_get_luma_subsampling_422_lbd(tx_size);
353 }
354 return cfl_get_luma_subsampling_444_lbd(tx_size);
355 }
356
cfl_store(CFL_CTX * cfl,const uint8_t * input,int input_stride,int row,int col,TX_SIZE tx_size,int use_hbd)357 static void cfl_store(CFL_CTX *cfl, const uint8_t *input, int input_stride,
358 int row, int col, TX_SIZE tx_size, int use_hbd) {
359 const int width = tx_size_wide[tx_size];
360 const int height = tx_size_high[tx_size];
361 const int tx_off_log2 = tx_size_wide_log2[0];
362 const int sub_x = cfl->subsampling_x;
363 const int sub_y = cfl->subsampling_y;
364 const int store_row = row << (tx_off_log2 - sub_y);
365 const int store_col = col << (tx_off_log2 - sub_x);
366 const int store_height = height >> sub_y;
367 const int store_width = width >> sub_x;
368
369 // Invalidate current parameters
370 cfl->are_parameters_computed = 0;
371
372 // Store the surface of the pixel buffer that was written to, this way we
373 // can manage chroma overrun (e.g. when the chroma surfaces goes beyond the
374 // frame boundary)
375 if (col == 0 && row == 0) {
376 cfl->buf_width = store_width;
377 cfl->buf_height = store_height;
378 } else {
379 cfl->buf_width = OD_MAXI(store_col + store_width, cfl->buf_width);
380 cfl->buf_height = OD_MAXI(store_row + store_height, cfl->buf_height);
381 }
382
383 // Check that we will remain inside the pixel buffer.
384 assert(store_row + store_height <= CFL_BUF_LINE);
385 assert(store_col + store_width <= CFL_BUF_LINE);
386
387 // Store the input into the CfL pixel buffer
388 uint16_t *recon_buf_q3 =
389 cfl->recon_buf_q3 + (store_row * CFL_BUF_LINE + store_col);
390
391 if (use_hbd) {
392 cfl_subsampling_hbd(tx_size, sub_x, sub_y)(CONVERT_TO_SHORTPTR(input),
393 input_stride, recon_buf_q3);
394 } else {
395 cfl_subsampling_lbd(tx_size, sub_x, sub_y)(input, input_stride,
396 recon_buf_q3);
397 }
398 }
399
400 // Adjust the row and column of blocks smaller than 8X8, as chroma-referenced
401 // and non-chroma-referenced blocks are stored together in the CfL buffer.
sub8x8_adjust_offset(const CFL_CTX * cfl,int * row_out,int * col_out)402 static INLINE void sub8x8_adjust_offset(const CFL_CTX *cfl, int *row_out,
403 int *col_out) {
404 // Increment row index for bottom: 8x4, 16x4 or both bottom 4x4s.
405 if ((cfl->mi_row & 0x01) && cfl->subsampling_y) {
406 assert(*row_out == 0);
407 (*row_out)++;
408 }
409
410 // Increment col index for right: 4x8, 4x16 or both right 4x4s.
411 if ((cfl->mi_col & 0x01) && cfl->subsampling_x) {
412 assert(*col_out == 0);
413 (*col_out)++;
414 }
415 }
416
cfl_store_tx(MACROBLOCKD * const xd,int row,int col,TX_SIZE tx_size,BLOCK_SIZE bsize)417 void cfl_store_tx(MACROBLOCKD *const xd, int row, int col, TX_SIZE tx_size,
418 BLOCK_SIZE bsize) {
419 CFL_CTX *const cfl = &xd->cfl;
420 struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
421 uint8_t *dst =
422 &pd->dst.buf[(row * pd->dst.stride + col) << tx_size_wide_log2[0]];
423
424 if (block_size_high[bsize] == 4 || block_size_wide[bsize] == 4) {
425 // Only dimensions of size 4 can have an odd offset.
426 assert(!((col & 1) && tx_size_wide[tx_size] != 4));
427 assert(!((row & 1) && tx_size_high[tx_size] != 4));
428 sub8x8_adjust_offset(cfl, &row, &col);
429 }
430 cfl_store(cfl, dst, pd->dst.stride, row, col, tx_size,
431 get_bitdepth_data_path_index(xd));
432 }
433
cfl_store_block(MACROBLOCKD * const xd,BLOCK_SIZE bsize,TX_SIZE tx_size)434 void cfl_store_block(MACROBLOCKD *const xd, BLOCK_SIZE bsize, TX_SIZE tx_size) {
435 CFL_CTX *const cfl = &xd->cfl;
436 struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
437 int row = 0;
438 int col = 0;
439
440 if (block_size_high[bsize] == 4 || block_size_wide[bsize] == 4) {
441 sub8x8_adjust_offset(cfl, &row, &col);
442 }
443 const int width = max_intra_block_width(xd, bsize, AOM_PLANE_Y, tx_size);
444 const int height = max_intra_block_height(xd, bsize, AOM_PLANE_Y, tx_size);
445 tx_size = get_tx_size(width, height);
446 cfl_store(cfl, pd->dst.buf, pd->dst.stride, row, col, tx_size,
447 get_bitdepth_data_path_index(xd));
448 }
449