1 /*
2 * Copyright (c) 2023, 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 "config/aom_dsp_rtcd.h"
13 #include "config/av1_rtcd.h"
14
15 #include "av1/common/reconinter.h"
16
17 #include "av1/encoder/encodemv.h"
18 #include "av1/encoder/nonrd_opt.h"
19 #include "av1/encoder/rdopt.h"
20
21 static const SCAN_ORDER av1_fast_idtx_scan_order_16x16 = {
22 av1_fast_idtx_scan_16x16, av1_fast_idtx_iscan_16x16
23 };
24
25 #define DECLARE_BLOCK_YRD_BUFFERS() \
26 DECLARE_ALIGNED(64, tran_low_t, dqcoeff_buf[16 * 16]); \
27 DECLARE_ALIGNED(64, tran_low_t, qcoeff_buf[16 * 16]); \
28 DECLARE_ALIGNED(64, tran_low_t, coeff_buf[16 * 16]); \
29 uint16_t eob[1];
30
31 #define DECLARE_BLOCK_YRD_VARS() \
32 /* When is_tx_8x8_dual_applicable is true, we compute the txfm for the \
33 * entire bsize and write macroblock_plane::coeff. So low_coeff is kept \
34 * as a non-const so we can reassign it to macroblock_plane::coeff. */ \
35 int16_t *low_coeff = (int16_t *)coeff_buf; \
36 int16_t *const low_qcoeff = (int16_t *)qcoeff_buf; \
37 int16_t *const low_dqcoeff = (int16_t *)dqcoeff_buf; \
38 const int diff_stride = bw;
39
40 #define DECLARE_LOOP_VARS_BLOCK_YRD() \
41 const int16_t *src_diff = &p->src_diff[(r * diff_stride + c) << 2];
42
update_yrd_loop_vars(MACROBLOCK * x,int * skippable,int step,int ncoeffs,int16_t * const low_coeff,int16_t * const low_qcoeff,int16_t * const low_dqcoeff,RD_STATS * this_rdc,int * eob_cost,int tx_blk_id)43 static AOM_FORCE_INLINE void update_yrd_loop_vars(
44 MACROBLOCK *x, int *skippable, int step, int ncoeffs,
45 int16_t *const low_coeff, int16_t *const low_qcoeff,
46 int16_t *const low_dqcoeff, RD_STATS *this_rdc, int *eob_cost,
47 int tx_blk_id) {
48 const int is_txfm_skip = (ncoeffs == 0);
49 *skippable &= is_txfm_skip;
50 x->txfm_search_info.blk_skip[tx_blk_id] = is_txfm_skip;
51 *eob_cost += get_msb(ncoeffs + 1);
52 if (ncoeffs == 1)
53 this_rdc->rate += (int)abs(low_qcoeff[0]);
54 else if (ncoeffs > 1)
55 this_rdc->rate += aom_satd_lp(low_qcoeff, step << 4);
56
57 this_rdc->dist += av1_block_error_lp(low_coeff, low_dqcoeff, step << 4) >> 2;
58 }
59
aom_process_hadamard_lp_8x16(MACROBLOCK * x,int max_blocks_high,int max_blocks_wide,int num_4x4_w,int step,int block_step)60 static inline void aom_process_hadamard_lp_8x16(MACROBLOCK *x,
61 int max_blocks_high,
62 int max_blocks_wide,
63 int num_4x4_w, int step,
64 int block_step) {
65 struct macroblock_plane *const p = &x->plane[AOM_PLANE_Y];
66 const int bw = 4 * num_4x4_w;
67 const int num_4x4 = AOMMIN(num_4x4_w, max_blocks_wide);
68 int block = 0;
69
70 for (int r = 0; r < max_blocks_high; r += block_step) {
71 for (int c = 0; c < num_4x4; c += 2 * block_step) {
72 const int16_t *src_diff = &p->src_diff[(r * bw + c) << 2];
73 int16_t *low_coeff = (int16_t *)p->coeff + BLOCK_OFFSET(block);
74 aom_hadamard_lp_8x8_dual(src_diff, (ptrdiff_t)bw, low_coeff);
75 block += 2 * step;
76 }
77 }
78 }
79
80 #if CONFIG_AV1_HIGHBITDEPTH
81 #define DECLARE_BLOCK_YRD_HBD_VARS() \
82 tran_low_t *const coeff = coeff_buf; \
83 tran_low_t *const qcoeff = qcoeff_buf; \
84 tran_low_t *const dqcoeff = dqcoeff_buf;
85
update_yrd_loop_vars_hbd(MACROBLOCK * x,int * skippable,int step,int ncoeffs,tran_low_t * const coeff,tran_low_t * const qcoeff,tran_low_t * const dqcoeff,RD_STATS * this_rdc,int * eob_cost,int tx_blk_id)86 static AOM_FORCE_INLINE void update_yrd_loop_vars_hbd(
87 MACROBLOCK *x, int *skippable, int step, int ncoeffs,
88 tran_low_t *const coeff, tran_low_t *const qcoeff,
89 tran_low_t *const dqcoeff, RD_STATS *this_rdc, int *eob_cost,
90 int tx_blk_id) {
91 const MACROBLOCKD *xd = &x->e_mbd;
92 const int is_txfm_skip = (ncoeffs == 0);
93 *skippable &= is_txfm_skip;
94 x->txfm_search_info.blk_skip[tx_blk_id] = is_txfm_skip;
95 *eob_cost += get_msb(ncoeffs + 1);
96
97 int64_t dummy;
98 if (ncoeffs == 1)
99 this_rdc->rate += (int)abs(qcoeff[0]);
100 else if (ncoeffs > 1)
101 this_rdc->rate += aom_satd(qcoeff, step << 4);
102 this_rdc->dist +=
103 av1_highbd_block_error(coeff, dqcoeff, step << 4, &dummy, xd->bd) >> 2;
104 }
105 #endif
106
107 /*!\brief Calculates RD Cost using Hadamard transform.
108 *
109 * \ingroup nonrd_mode_search
110 * \callgraph
111 * \callergraph
112 * Calculates RD Cost using Hadamard transform. For low bit depth this function
113 * uses low-precision set of functions (16-bit) and 32 bit for high bit depth
114 * \param[in] x Pointer to structure holding all the data for
115 the current macroblock
116 * \param[in] this_rdc Pointer to calculated RD Cost
117 * \param[in] skippable Pointer to a flag indicating possible tx skip
118 * \param[in] bsize Current block size
119 * \param[in] tx_size Transform size
120 * \param[in] is_inter_mode Flag to indicate inter mode
121 *
122 * \remark Nothing is returned. Instead, calculated RD cost is placed to
123 * \c this_rdc. \c skippable flag is set if there is no non-zero quantized
124 * coefficients for Hadamard transform
125 */
av1_block_yrd(MACROBLOCK * x,RD_STATS * this_rdc,int * skippable,BLOCK_SIZE bsize,TX_SIZE tx_size)126 void av1_block_yrd(MACROBLOCK *x, RD_STATS *this_rdc, int *skippable,
127 BLOCK_SIZE bsize, TX_SIZE tx_size) {
128 MACROBLOCKD *xd = &x->e_mbd;
129 const struct macroblockd_plane *pd = &xd->plane[AOM_PLANE_Y];
130 struct macroblock_plane *const p = &x->plane[AOM_PLANE_Y];
131 assert(bsize < BLOCK_SIZES_ALL);
132 const int num_4x4_w = mi_size_wide[bsize];
133 const int num_4x4_h = mi_size_high[bsize];
134 const int step = 1 << (tx_size << 1);
135 const int block_step = (1 << tx_size);
136 const int row_step = step * num_4x4_w >> tx_size;
137 int block = 0;
138 const int max_blocks_wide =
139 num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 : xd->mb_to_right_edge >> 5);
140 const int max_blocks_high =
141 num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 : xd->mb_to_bottom_edge >> 5);
142 int eob_cost = 0;
143 const int bw = 4 * num_4x4_w;
144 const int bh = 4 * num_4x4_h;
145 const int use_hbd = is_cur_buf_hbd(xd);
146 int num_blk_skip_w = num_4x4_w;
147
148 #if CONFIG_AV1_HIGHBITDEPTH
149 if (use_hbd) {
150 aom_highbd_subtract_block(bh, bw, p->src_diff, bw, p->src.buf,
151 p->src.stride, pd->dst.buf, pd->dst.stride);
152 } else {
153 aom_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
154 pd->dst.buf, pd->dst.stride);
155 }
156 #else
157 aom_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
158 pd->dst.buf, pd->dst.stride);
159 #endif
160
161 // Keep the intermediate value on the stack here. Writing directly to
162 // skippable causes speed regression due to load-and-store issues in
163 // update_yrd_loop_vars.
164 int temp_skippable = 1;
165 this_rdc->dist = 0;
166 this_rdc->rate = 0;
167 // For block sizes 8x16 or above, Hadamard txfm of two adjacent 8x8 blocks
168 // can be done per function call. Hence the call of Hadamard txfm is
169 // abstracted here for the specified cases.
170 int is_tx_8x8_dual_applicable =
171 (tx_size == TX_8X8 && block_size_wide[bsize] >= 16 &&
172 block_size_high[bsize] >= 8);
173
174 #if CONFIG_AV1_HIGHBITDEPTH
175 // As of now, dual implementation of hadamard txfm is available for low
176 // bitdepth.
177 if (use_hbd) is_tx_8x8_dual_applicable = 0;
178 #endif
179
180 if (is_tx_8x8_dual_applicable) {
181 aom_process_hadamard_lp_8x16(x, max_blocks_high, max_blocks_wide, num_4x4_w,
182 step, block_step);
183 }
184
185 const SCAN_ORDER *const scan_order = &av1_scan_orders[tx_size][DCT_DCT];
186 DECLARE_BLOCK_YRD_BUFFERS()
187 DECLARE_BLOCK_YRD_VARS()
188 #if CONFIG_AV1_HIGHBITDEPTH
189 DECLARE_BLOCK_YRD_HBD_VARS()
190 #else
191 (void)use_hbd;
192 #endif
193
194 // Keep track of the row and column of the blocks we use so that we know
195 // if we are in the unrestricted motion border.
196 for (int r = 0; r < max_blocks_high; r += block_step) {
197 for (int c = 0, s = 0; c < max_blocks_wide; c += block_step, s += step) {
198 DECLARE_LOOP_VARS_BLOCK_YRD()
199
200 switch (tx_size) {
201 #if CONFIG_AV1_HIGHBITDEPTH
202 case TX_16X16:
203 if (use_hbd) {
204 aom_hadamard_16x16(src_diff, diff_stride, coeff);
205 av1_quantize_fp(coeff, 16 * 16, p->zbin_QTX, p->round_fp_QTX,
206 p->quant_fp_QTX, p->quant_shift_QTX, qcoeff,
207 dqcoeff, p->dequant_QTX, eob,
208 // default_scan_fp_16x16_transpose and
209 // av1_default_iscan_fp_16x16_transpose have to be
210 // used together.
211 default_scan_fp_16x16_transpose,
212 av1_default_iscan_fp_16x16_transpose);
213 } else {
214 aom_hadamard_lp_16x16(src_diff, diff_stride, low_coeff);
215 av1_quantize_lp(low_coeff, 16 * 16, p->round_fp_QTX,
216 p->quant_fp_QTX, low_qcoeff, low_dqcoeff,
217 p->dequant_QTX, eob,
218 // default_scan_lp_16x16_transpose and
219 // av1_default_iscan_lp_16x16_transpose have to be
220 // used together.
221 default_scan_lp_16x16_transpose,
222 av1_default_iscan_lp_16x16_transpose);
223 }
224 break;
225 case TX_8X8:
226 if (use_hbd) {
227 aom_hadamard_8x8(src_diff, diff_stride, coeff);
228 av1_quantize_fp(
229 coeff, 8 * 8, p->zbin_QTX, p->round_fp_QTX, p->quant_fp_QTX,
230 p->quant_shift_QTX, qcoeff, dqcoeff, p->dequant_QTX, eob,
231 default_scan_8x8_transpose, av1_default_iscan_8x8_transpose);
232 } else {
233 if (is_tx_8x8_dual_applicable) {
234 // The coeffs are pre-computed for the whole block, so re-assign
235 // low_coeff to the appropriate location.
236 const int block_offset = BLOCK_OFFSET(block + s);
237 low_coeff = (int16_t *)p->coeff + block_offset;
238 } else {
239 aom_hadamard_lp_8x8(src_diff, diff_stride, low_coeff);
240 }
241 av1_quantize_lp(
242 low_coeff, 8 * 8, p->round_fp_QTX, p->quant_fp_QTX, low_qcoeff,
243 low_dqcoeff, p->dequant_QTX, eob,
244 // default_scan_8x8_transpose and
245 // av1_default_iscan_8x8_transpose have to be used together.
246 default_scan_8x8_transpose, av1_default_iscan_8x8_transpose);
247 }
248 break;
249 default:
250 assert(tx_size == TX_4X4);
251 // In tx_size=4x4 case, aom_fdct4x4 and aom_fdct4x4_lp generate
252 // normal coefficients order, so we don't need to change the scan
253 // order here.
254 if (use_hbd) {
255 aom_fdct4x4(src_diff, coeff, diff_stride);
256 av1_quantize_fp(coeff, 4 * 4, p->zbin_QTX, p->round_fp_QTX,
257 p->quant_fp_QTX, p->quant_shift_QTX, qcoeff,
258 dqcoeff, p->dequant_QTX, eob, scan_order->scan,
259 scan_order->iscan);
260 } else {
261 aom_fdct4x4_lp(src_diff, low_coeff, diff_stride);
262 av1_quantize_lp(low_coeff, 4 * 4, p->round_fp_QTX, p->quant_fp_QTX,
263 low_qcoeff, low_dqcoeff, p->dequant_QTX, eob,
264 scan_order->scan, scan_order->iscan);
265 }
266 break;
267 #else
268 case TX_16X16:
269 aom_hadamard_lp_16x16(src_diff, diff_stride, low_coeff);
270 av1_quantize_lp(low_coeff, 16 * 16, p->round_fp_QTX, p->quant_fp_QTX,
271 low_qcoeff, low_dqcoeff, p->dequant_QTX, eob,
272 default_scan_lp_16x16_transpose,
273 av1_default_iscan_lp_16x16_transpose);
274 break;
275 case TX_8X8:
276 if (is_tx_8x8_dual_applicable) {
277 // The coeffs are pre-computed for the whole block, so re-assign
278 // low_coeff to the appropriate location.
279 const int block_offset = BLOCK_OFFSET(block + s);
280 low_coeff = (int16_t *)p->coeff + block_offset;
281 } else {
282 aom_hadamard_lp_8x8(src_diff, diff_stride, low_coeff);
283 }
284 av1_quantize_lp(low_coeff, 8 * 8, p->round_fp_QTX, p->quant_fp_QTX,
285 low_qcoeff, low_dqcoeff, p->dequant_QTX, eob,
286 default_scan_8x8_transpose,
287 av1_default_iscan_8x8_transpose);
288 break;
289 default:
290 aom_fdct4x4_lp(src_diff, low_coeff, diff_stride);
291 av1_quantize_lp(low_coeff, 4 * 4, p->round_fp_QTX, p->quant_fp_QTX,
292 low_qcoeff, low_dqcoeff, p->dequant_QTX, eob,
293 scan_order->scan, scan_order->iscan);
294 break;
295 #endif
296 }
297 assert(*eob <= 1024);
298 #if CONFIG_AV1_HIGHBITDEPTH
299 if (use_hbd)
300 update_yrd_loop_vars_hbd(x, &temp_skippable, step, *eob, coeff, qcoeff,
301 dqcoeff, this_rdc, &eob_cost,
302 r * num_blk_skip_w + c);
303 else
304 #endif
305 update_yrd_loop_vars(x, &temp_skippable, step, *eob, low_coeff,
306 low_qcoeff, low_dqcoeff, this_rdc, &eob_cost,
307 r * num_blk_skip_w + c);
308 }
309 block += row_step;
310 }
311
312 this_rdc->skip_txfm = *skippable = temp_skippable;
313 if (this_rdc->sse < INT64_MAX) {
314 this_rdc->sse = (this_rdc->sse << 6) >> 2;
315 if (temp_skippable) {
316 this_rdc->dist = 0;
317 this_rdc->dist = this_rdc->sse;
318 return;
319 }
320 }
321
322 // If skippable is set, rate gets clobbered later.
323 this_rdc->rate <<= (2 + AV1_PROB_COST_SHIFT);
324 this_rdc->rate += (eob_cost << AV1_PROB_COST_SHIFT);
325 }
326
327 // Explicitly enumerate the cases so the compiler can generate SIMD for the
328 // function. According to the disassembler, gcc generates SSE codes for each of
329 // the possible block sizes. The hottest case is tx_width 16, which takes up
330 // about 8% of the self cycle of av1_nonrd_pick_inter_mode_sb. Since
331 // av1_nonrd_pick_inter_mode_sb takes up about 3% of total encoding time, the
332 // potential room of improvement for writing AVX2 optimization is only 3% * 8% =
333 // 0.24% of total encoding time.
scale_square_buf_vals(int16_t * dst,int tx_width,const int16_t * src,int src_stride)334 static inline void scale_square_buf_vals(int16_t *dst, int tx_width,
335 const int16_t *src, int src_stride) {
336 #define DO_SCALING \
337 do { \
338 for (int idy = 0; idy < tx_width; ++idy) { \
339 for (int idx = 0; idx < tx_width; ++idx) { \
340 dst[idy * tx_width + idx] = src[idy * src_stride + idx] * 8; \
341 } \
342 } \
343 } while (0)
344
345 if (tx_width == 4) {
346 DO_SCALING;
347 } else if (tx_width == 8) {
348 DO_SCALING;
349 } else if (tx_width == 16) {
350 DO_SCALING;
351 } else {
352 assert(0);
353 }
354
355 #undef DO_SCALING
356 }
357
358 /*!\brief Calculates RD Cost when the block uses Identity transform.
359 * Note that this function is only for low bit depth encoding, since it
360 * is called in real-time mode for now, which sets high bit depth to 0:
361 * -DCONFIG_AV1_HIGHBITDEPTH=0
362 *
363 * \ingroup nonrd_mode_search
364 * \callgraph
365 * \callergraph
366 * Calculates RD Cost. For low bit depth this function
367 * uses low-precision set of functions (16-bit) and 32 bit for high bit depth
368 * \param[in] x Pointer to structure holding all the data for
369 the current macroblock
370 * \param[in] pred_buf Pointer to the prediction buffer
371 * \param[in] pred_stride Stride for the prediction buffer
372 * \param[in] this_rdc Pointer to calculated RD Cost
373 * \param[in] skippable Pointer to a flag indicating possible tx skip
374 * \param[in] bsize Current block size
375 * \param[in] tx_size Transform size
376 *
377 * \remark Nothing is returned. Instead, calculated RD cost is placed to
378 * \c this_rdc. \c skippable flag is set if all coefficients are zero.
379 */
av1_block_yrd_idtx(MACROBLOCK * x,const uint8_t * const pred_buf,int pred_stride,RD_STATS * this_rdc,int * skippable,BLOCK_SIZE bsize,TX_SIZE tx_size)380 void av1_block_yrd_idtx(MACROBLOCK *x, const uint8_t *const pred_buf,
381 int pred_stride, RD_STATS *this_rdc, int *skippable,
382 BLOCK_SIZE bsize, TX_SIZE tx_size) {
383 MACROBLOCKD *xd = &x->e_mbd;
384 struct macroblock_plane *const p = &x->plane[AOM_PLANE_Y];
385 assert(bsize < BLOCK_SIZES_ALL);
386 const int num_4x4_w = mi_size_wide[bsize];
387 const int num_4x4_h = mi_size_high[bsize];
388 const int step = 1 << (tx_size << 1);
389 const int block_step = (1 << tx_size);
390 const int max_blocks_wide =
391 num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 : xd->mb_to_right_edge >> 5);
392 const int max_blocks_high =
393 num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 : xd->mb_to_bottom_edge >> 5);
394 int eob_cost = 0;
395 const int bw = 4 * num_4x4_w;
396 const int bh = 4 * num_4x4_h;
397 const int num_blk_skip_w = num_4x4_w;
398 // Keep the intermediate value on the stack here. Writing directly to
399 // skippable causes speed regression due to load-and-store issues in
400 // update_yrd_loop_vars.
401 int temp_skippable = 1;
402 int tx_wd = 0;
403 const SCAN_ORDER *scan_order = NULL;
404 switch (tx_size) {
405 case TX_64X64:
406 assert(0); // Not implemented
407 break;
408 case TX_32X32:
409 assert(0); // Not used
410 break;
411 case TX_16X16:
412 scan_order = &av1_fast_idtx_scan_order_16x16;
413 tx_wd = 16;
414 break;
415 case TX_8X8:
416 scan_order = &av1_fast_idtx_scan_order_8x8;
417 tx_wd = 8;
418 break;
419 default:
420 assert(tx_size == TX_4X4);
421 scan_order = &av1_fast_idtx_scan_order_4x4;
422 tx_wd = 4;
423 break;
424 }
425 assert(scan_order != NULL);
426
427 this_rdc->dist = 0;
428 this_rdc->rate = 0;
429 aom_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
430 pred_buf, pred_stride);
431 // Keep track of the row and column of the blocks we use so that we know
432 // if we are in the unrestricted motion border.
433 DECLARE_BLOCK_YRD_BUFFERS()
434 DECLARE_BLOCK_YRD_VARS()
435 for (int r = 0; r < max_blocks_high; r += block_step) {
436 for (int c = 0, s = 0; c < max_blocks_wide; c += block_step, s += step) {
437 DECLARE_LOOP_VARS_BLOCK_YRD()
438 scale_square_buf_vals(low_coeff, tx_wd, src_diff, diff_stride);
439 av1_quantize_lp(low_coeff, tx_wd * tx_wd, p->round_fp_QTX,
440 p->quant_fp_QTX, low_qcoeff, low_dqcoeff, p->dequant_QTX,
441 eob, scan_order->scan, scan_order->iscan);
442 assert(*eob <= 1024);
443 update_yrd_loop_vars(x, &temp_skippable, step, *eob, low_coeff,
444 low_qcoeff, low_dqcoeff, this_rdc, &eob_cost,
445 r * num_blk_skip_w + c);
446 }
447 }
448 this_rdc->skip_txfm = *skippable = temp_skippable;
449 if (this_rdc->sse < INT64_MAX) {
450 this_rdc->sse = (this_rdc->sse << 6) >> 2;
451 if (temp_skippable) {
452 this_rdc->dist = 0;
453 this_rdc->dist = this_rdc->sse;
454 return;
455 }
456 }
457 // If skippable is set, rate gets clobbered later.
458 this_rdc->rate <<= (2 + AV1_PROB_COST_SHIFT);
459 this_rdc->rate += (eob_cost << AV1_PROB_COST_SHIFT);
460 }
461
av1_model_rd_for_sb_uv(AV1_COMP * cpi,BLOCK_SIZE plane_bsize,MACROBLOCK * x,MACROBLOCKD * xd,RD_STATS * this_rdc,int start_plane,int stop_plane)462 int64_t av1_model_rd_for_sb_uv(AV1_COMP *cpi, BLOCK_SIZE plane_bsize,
463 MACROBLOCK *x, MACROBLOCKD *xd,
464 RD_STATS *this_rdc, int start_plane,
465 int stop_plane) {
466 // Note our transform coeffs are 8 times an orthogonal transform.
467 // Hence quantizer step is also 8 times. To get effective quantizer
468 // we need to divide by 8 before sending to modeling function.
469 unsigned int sse;
470 int rate;
471 int64_t dist;
472 int plane;
473 int64_t tot_sse = 0;
474
475 this_rdc->rate = 0;
476 this_rdc->dist = 0;
477 this_rdc->skip_txfm = 0;
478
479 for (plane = start_plane; plane <= stop_plane; ++plane) {
480 struct macroblock_plane *const p = &x->plane[plane];
481 struct macroblockd_plane *const pd = &xd->plane[plane];
482 const uint32_t dc_quant = p->dequant_QTX[0];
483 const uint32_t ac_quant = p->dequant_QTX[1];
484 const BLOCK_SIZE bs = plane_bsize;
485 unsigned int var;
486 if (!x->color_sensitivity[COLOR_SENS_IDX(plane)]) continue;
487
488 var = cpi->ppi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf,
489 pd->dst.stride, &sse);
490 assert(sse >= var);
491 tot_sse += sse;
492
493 av1_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
494 dc_quant >> 3, &rate, &dist);
495
496 this_rdc->rate += rate >> 1;
497 this_rdc->dist += dist << 3;
498
499 av1_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs], ac_quant >> 3,
500 &rate, &dist);
501
502 this_rdc->rate += rate;
503 this_rdc->dist += dist << 4;
504 }
505
506 if (this_rdc->rate == 0) {
507 this_rdc->skip_txfm = 1;
508 }
509
510 if (RDCOST(x->rdmult, this_rdc->rate, this_rdc->dist) >=
511 RDCOST(x->rdmult, 0, tot_sse << 4)) {
512 this_rdc->rate = 0;
513 this_rdc->dist = tot_sse << 4;
514 this_rdc->skip_txfm = 1;
515 }
516
517 return tot_sse;
518 }
519
compute_intra_yprediction(const AV1_COMMON * cm,PREDICTION_MODE mode,BLOCK_SIZE bsize,MACROBLOCK * x,MACROBLOCKD * xd)520 static void compute_intra_yprediction(const AV1_COMMON *cm,
521 PREDICTION_MODE mode, BLOCK_SIZE bsize,
522 MACROBLOCK *x, MACROBLOCKD *xd) {
523 const SequenceHeader *seq_params = cm->seq_params;
524 struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
525 struct macroblock_plane *const p = &x->plane[AOM_PLANE_Y];
526 uint8_t *const src_buf_base = p->src.buf;
527 uint8_t *const dst_buf_base = pd->dst.buf;
528 const int src_stride = p->src.stride;
529 const int dst_stride = pd->dst.stride;
530 int plane = 0;
531 int row, col;
532 // block and transform sizes, in number of 4x4 blocks log 2 ("*_b")
533 // 4x4=0, 8x8=2, 16x16=4, 32x32=6, 64x64=8
534 // transform size varies per plane, look it up in a common way.
535 const TX_SIZE tx_size = max_txsize_lookup[bsize];
536 const BLOCK_SIZE plane_bsize =
537 get_plane_block_size(bsize, pd->subsampling_x, pd->subsampling_y);
538 // If mb_to_right_edge is < 0 we are in a situation in which
539 // the current block size extends into the UMV and we won't
540 // visit the sub blocks that are wholly within the UMV.
541 const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
542 const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
543 // Keep track of the row and column of the blocks we use so that we know
544 // if we are in the unrestricted motion border.
545 for (row = 0; row < max_blocks_high; row += (1 << tx_size)) {
546 // Skip visiting the sub blocks that are wholly within the UMV.
547 for (col = 0; col < max_blocks_wide; col += (1 << tx_size)) {
548 p->src.buf = &src_buf_base[4 * (row * (int64_t)src_stride + col)];
549 pd->dst.buf = &dst_buf_base[4 * (row * (int64_t)dst_stride + col)];
550 av1_predict_intra_block(
551 xd, seq_params->sb_size, seq_params->enable_intra_edge_filter,
552 block_size_wide[bsize], block_size_high[bsize], tx_size, mode, 0, 0,
553 FILTER_INTRA_MODES, pd->dst.buf, dst_stride, pd->dst.buf, dst_stride,
554 0, 0, plane);
555 }
556 }
557 p->src.buf = src_buf_base;
558 pd->dst.buf = dst_buf_base;
559 }
560
561 // Checks whether Intra mode needs to be pruned based on
562 // 'intra_y_mode_bsize_mask_nrd' and 'prune_hv_pred_modes_using_blksad'
563 // speed features.
is_prune_intra_mode(AV1_COMP * cpi,int mode_index,int force_intra_check,BLOCK_SIZE bsize,uint8_t segment_id,SOURCE_SAD source_sad_nonrd,uint8_t color_sensitivity[MAX_MB_PLANE-1])564 static inline bool is_prune_intra_mode(
565 AV1_COMP *cpi, int mode_index, int force_intra_check, BLOCK_SIZE bsize,
566 uint8_t segment_id, SOURCE_SAD source_sad_nonrd,
567 uint8_t color_sensitivity[MAX_MB_PLANE - 1]) {
568 const PREDICTION_MODE this_mode = intra_mode_list[mode_index];
569 if (mode_index > 2 || force_intra_check == 0) {
570 if (!((1 << this_mode) & cpi->sf.rt_sf.intra_y_mode_bsize_mask_nrd[bsize]))
571 return true;
572
573 if (this_mode == DC_PRED) return false;
574
575 if (!cpi->sf.rt_sf.prune_hv_pred_modes_using_src_sad) return false;
576
577 const bool has_color_sensitivity =
578 color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] &&
579 color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)];
580 if (has_color_sensitivity &&
581 (cpi->rc.frame_source_sad > 1.1 * cpi->rc.avg_source_sad ||
582 cyclic_refresh_segment_id_boosted(segment_id) ||
583 source_sad_nonrd > kMedSad))
584 return false;
585
586 return true;
587 }
588 return false;
589 }
590
591 /*!\brief Estimation of RD cost of an intra mode for Non-RD optimized case.
592 *
593 * \ingroup nonrd_mode_search
594 * \callgraph
595 * \callergraph
596 * Calculates RD Cost for an intra mode for a single TX block using Hadamard
597 * transform.
598 * \param[in] plane Color plane
599 * \param[in] block Index of a TX block in a prediction block
600 * \param[in] row Row of a current TX block
601 * \param[in] col Column of a current TX block
602 * \param[in] plane_bsize Block size of a current prediction block
603 * \param[in] tx_size Transform size
604 * \param[in] arg Pointer to a structure that holds parameters
605 * for intra mode search
606 *
607 * \remark Nothing is returned. Instead, best mode and RD Cost of the best mode
608 * are set in \c args->rdc and \c args->mode
609 */
av1_estimate_block_intra(int plane,int block,int row,int col,BLOCK_SIZE plane_bsize,TX_SIZE tx_size,void * arg)610 void av1_estimate_block_intra(int plane, int block, int row, int col,
611 BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
612 void *arg) {
613 struct estimate_block_intra_args *const args = arg;
614 AV1_COMP *const cpi = args->cpi;
615 AV1_COMMON *const cm = &cpi->common;
616 MACROBLOCK *const x = args->x;
617 MACROBLOCKD *const xd = &x->e_mbd;
618 struct macroblock_plane *const p = &x->plane[plane];
619 struct macroblockd_plane *const pd = &xd->plane[plane];
620 const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
621 uint8_t *const src_buf_base = p->src.buf;
622 uint8_t *const dst_buf_base = pd->dst.buf;
623 const int64_t src_stride = p->src.stride;
624 const int64_t dst_stride = pd->dst.stride;
625
626 (void)block;
627
628 av1_predict_intra_block_facade(cm, xd, plane, col, row, tx_size);
629
630 if (args->prune_mode_based_on_sad) {
631 unsigned int this_sad = cpi->ppi->fn_ptr[plane_bsize].sdf(
632 p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride);
633 const unsigned int sad_threshold =
634 args->best_sad != UINT_MAX ? args->best_sad + (args->best_sad >> 4)
635 : UINT_MAX;
636 // Skip the evaluation of current mode if its SAD is more than a threshold.
637 if (this_sad > sad_threshold) {
638 // For the current mode, set rate and distortion to maximum possible
639 // values and return.
640 // Note: args->rdc->rate is checked in av1_nonrd_pick_intra_mode() to skip
641 // the evaluation of the current mode.
642 args->rdc->rate = INT_MAX;
643 args->rdc->dist = INT64_MAX;
644 return;
645 }
646 if (this_sad < args->best_sad) {
647 args->best_sad = this_sad;
648 }
649 }
650
651 RD_STATS this_rdc;
652 av1_invalid_rd_stats(&this_rdc);
653
654 p->src.buf = &src_buf_base[4 * (row * src_stride + col)];
655 pd->dst.buf = &dst_buf_base[4 * (row * dst_stride + col)];
656
657 if (plane == 0) {
658 av1_block_yrd(x, &this_rdc, &args->skippable, bsize_tx,
659 AOMMIN(tx_size, TX_16X16));
660 } else {
661 av1_model_rd_for_sb_uv(cpi, bsize_tx, x, xd, &this_rdc, plane, plane);
662 }
663
664 p->src.buf = src_buf_base;
665 pd->dst.buf = dst_buf_base;
666 assert(args->rdc->rate != INT_MAX && args->rdc->dist != INT64_MAX);
667 args->rdc->rate += this_rdc.rate;
668 args->rdc->dist += this_rdc.dist;
669 }
670
671 /*!\brief Estimates best intra mode for inter mode search
672 *
673 * \ingroup nonrd_mode_search
674 * \callgraph
675 * \callergraph
676 *
677 * Using heuristics based on best inter mode, block size, and other decides
678 * whether to check intra modes. If so, estimates and selects best intra mode
679 * from the reduced set of intra modes (max 4 intra modes checked)
680 *
681 * \param[in] cpi Top-level encoder structure
682 * \param[in] x Pointer to structure holding all the
683 * data for the current macroblock
684 * \param[in] bsize Current block size
685 * \param[in] best_early_term Flag, indicating that TX for the
686 * best inter mode was skipped
687 * \param[in] ref_cost_intra Cost of signalling intra mode
688 * \param[in] reuse_prediction Flag, indicating prediction re-use
689 * \param[in] orig_dst Original destination buffer
690 * \param[in] tmp_buffers Pointer to a temporary buffers for
691 * prediction re-use
692 * \param[out] this_mode_pred Pointer to store prediction buffer
693 * for prediction re-use
694 * \param[in] best_rdc Pointer to RD cost for the best
695 * selected intra mode
696 * \param[in] best_pickmode Pointer to a structure containing
697 * best mode picked so far
698 * \param[in] ctx Pointer to structure holding coding
699 * contexts and modes for the block
700 *
701 * \remark Nothing is returned. Instead, calculated RD cost is placed to
702 * \c best_rdc and best selected mode is placed to \c best_pickmode
703 *
704 */
av1_estimate_intra_mode(AV1_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int best_early_term,unsigned int ref_cost_intra,int reuse_prediction,struct buf_2d * orig_dst,PRED_BUFFER * tmp_buffers,PRED_BUFFER ** this_mode_pred,RD_STATS * best_rdc,BEST_PICKMODE * best_pickmode,PICK_MODE_CONTEXT * ctx)705 void av1_estimate_intra_mode(AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
706 int best_early_term, unsigned int ref_cost_intra,
707 int reuse_prediction, struct buf_2d *orig_dst,
708 PRED_BUFFER *tmp_buffers,
709 PRED_BUFFER **this_mode_pred, RD_STATS *best_rdc,
710 BEST_PICKMODE *best_pickmode,
711 PICK_MODE_CONTEXT *ctx) {
712 AV1_COMMON *const cm = &cpi->common;
713 MACROBLOCKD *const xd = &x->e_mbd;
714 MB_MODE_INFO *const mi = xd->mi[0];
715 const TxfmSearchParams *txfm_params = &x->txfm_search_params;
716 const unsigned char segment_id = mi->segment_id;
717 const int *const rd_threshes = cpi->rd.threshes[segment_id][bsize];
718 const int *const rd_thresh_freq_fact = x->thresh_freq_fact[bsize];
719 const bool is_screen_content =
720 cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN;
721 struct macroblockd_plane *const pd = &xd->plane[AOM_PLANE_Y];
722 const REAL_TIME_SPEED_FEATURES *const rt_sf = &cpi->sf.rt_sf;
723
724 const CommonQuantParams *quant_params = &cm->quant_params;
725
726 RD_STATS this_rdc;
727
728 int intra_cost_penalty = av1_get_intra_cost_penalty(
729 quant_params->base_qindex, quant_params->y_dc_delta_q,
730 cm->seq_params->bit_depth);
731 int64_t inter_mode_thresh =
732 RDCOST(x->rdmult, ref_cost_intra + intra_cost_penalty, 0);
733 int perform_intra_pred = rt_sf->check_intra_pred_nonrd;
734 int force_intra_check = 0;
735 // For spatial enhancement layer: turn off intra prediction if the
736 // previous spatial layer as golden ref is not chosen as best reference.
737 // only do this for temporal enhancement layer and on non-key frames.
738 if (cpi->svc.spatial_layer_id > 0 &&
739 best_pickmode->best_ref_frame != GOLDEN_FRAME &&
740 cpi->svc.temporal_layer_id > 0 &&
741 !cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame)
742 perform_intra_pred = 0;
743
744 int do_early_exit_rdthresh = 1;
745
746 uint32_t spatial_var_thresh = 50;
747 int motion_thresh = 32;
748 // Adjust thresholds to make intra mode likely tested if the other
749 // references (golden, alt) are skipped/not checked. For now always
750 // adjust for svc mode.
751 if (cpi->ppi->use_svc || (rt_sf->use_nonrd_altref_frame == 0 &&
752 rt_sf->nonrd_prune_ref_frame_search > 0)) {
753 spatial_var_thresh = 150;
754 motion_thresh = 0;
755 }
756
757 // Some adjustments to checking intra mode based on source variance.
758 if (x->source_variance < spatial_var_thresh) {
759 // If the best inter mode is large motion or non-LAST ref reduce intra cost
760 // penalty, so intra mode is more likely tested.
761 if (best_rdc->rdcost != INT64_MAX &&
762 (best_pickmode->best_ref_frame != LAST_FRAME ||
763 abs(mi->mv[0].as_mv.row) >= motion_thresh ||
764 abs(mi->mv[0].as_mv.col) >= motion_thresh)) {
765 intra_cost_penalty = intra_cost_penalty >> 2;
766 inter_mode_thresh =
767 RDCOST(x->rdmult, ref_cost_intra + intra_cost_penalty, 0);
768 do_early_exit_rdthresh = 0;
769 }
770 if ((x->source_variance < AOMMAX(50, (spatial_var_thresh >> 1)) &&
771 x->content_state_sb.source_sad_nonrd >= kHighSad) ||
772 (is_screen_content && x->source_variance < 50 &&
773 ((bsize >= BLOCK_32X32 &&
774 x->content_state_sb.source_sad_nonrd != kZeroSad) ||
775 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] == 1 ||
776 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] == 1)))
777 force_intra_check = 1;
778 // For big blocks worth checking intra (since only DC will be checked),
779 // even if best_early_term is set.
780 if (bsize >= BLOCK_32X32) best_early_term = 0;
781 } else if (rt_sf->source_metrics_sb_nonrd &&
782 x->content_state_sb.source_sad_nonrd <= kLowSad) {
783 perform_intra_pred = 0;
784 }
785
786 if (best_rdc->skip_txfm && best_pickmode->best_mode_initial_skip_flag) {
787 if (rt_sf->skip_intra_pred == 1 && best_pickmode->best_mode != NEWMV)
788 perform_intra_pred = 0;
789 else if (rt_sf->skip_intra_pred == 2)
790 perform_intra_pred = 0;
791 }
792
793 if (!(best_rdc->rdcost == INT64_MAX || force_intra_check ||
794 (perform_intra_pred && !best_early_term &&
795 bsize <= cpi->sf.part_sf.max_intra_bsize))) {
796 return;
797 }
798
799 // Early exit based on RD cost calculated using known rate. When
800 // is_screen_content is true, more bias is given to intra modes. Hence,
801 // considered conservative threshold in early exit for the same.
802 const int64_t known_rd = is_screen_content
803 ? CALC_BIASED_RDCOST(inter_mode_thresh)
804 : inter_mode_thresh;
805 if (known_rd > best_rdc->rdcost) return;
806
807 struct estimate_block_intra_args args;
808 init_estimate_block_intra_args(&args, cpi, x);
809 TX_SIZE intra_tx_size = AOMMIN(
810 AOMMIN(max_txsize_lookup[bsize],
811 tx_mode_to_biggest_tx_size[txfm_params->tx_mode_search_type]),
812 TX_16X16);
813 if (is_screen_content && cpi->rc.high_source_sad &&
814 x->source_variance > spatial_var_thresh && bsize <= BLOCK_16X16)
815 intra_tx_size = TX_4X4;
816
817 PRED_BUFFER *const best_pred = best_pickmode->best_pred;
818 if (reuse_prediction && best_pred != NULL) {
819 const int bh = block_size_high[bsize];
820 const int bw = block_size_wide[bsize];
821 if (best_pred->data == orig_dst->buf) {
822 *this_mode_pred = &tmp_buffers[get_pred_buffer(tmp_buffers, 3)];
823 aom_convolve_copy(best_pred->data, best_pred->stride,
824 (*this_mode_pred)->data, (*this_mode_pred)->stride, bw,
825 bh);
826 best_pickmode->best_pred = *this_mode_pred;
827 }
828 }
829 pd->dst = *orig_dst;
830
831 for (int midx = 0; midx < RTC_INTRA_MODES; ++midx) {
832 const PREDICTION_MODE this_mode = intra_mode_list[midx];
833 const THR_MODES mode_index = mode_idx[INTRA_FRAME][mode_offset(this_mode)];
834 const int64_t mode_rd_thresh = rd_threshes[mode_index];
835
836 if (is_prune_intra_mode(cpi, midx, force_intra_check, bsize, segment_id,
837 x->content_state_sb.source_sad_nonrd,
838 x->color_sensitivity))
839 continue;
840
841 if (is_screen_content && rt_sf->source_metrics_sb_nonrd) {
842 // For spatially flat blocks with zero motion only check
843 // DC mode.
844 if (x->content_state_sb.source_sad_nonrd == kZeroSad &&
845 x->source_variance == 0 && this_mode != DC_PRED)
846 continue;
847 // Only test Intra for big blocks if spatial_variance is small.
848 else if (bsize > BLOCK_32X32 && x->source_variance > 50)
849 continue;
850 }
851
852 if (rd_less_than_thresh(best_rdc->rdcost, mode_rd_thresh,
853 rd_thresh_freq_fact[mode_index]) &&
854 (do_early_exit_rdthresh || this_mode == SMOOTH_PRED)) {
855 continue;
856 }
857 const BLOCK_SIZE uv_bsize =
858 get_plane_block_size(bsize, xd->plane[AOM_PLANE_U].subsampling_x,
859 xd->plane[AOM_PLANE_U].subsampling_y);
860
861 mi->mode = this_mode;
862 mi->ref_frame[0] = INTRA_FRAME;
863 mi->ref_frame[1] = NONE_FRAME;
864
865 av1_invalid_rd_stats(&this_rdc);
866 args.mode = this_mode;
867 args.skippable = 1;
868 args.rdc = &this_rdc;
869 mi->tx_size = intra_tx_size;
870 compute_intra_yprediction(cm, this_mode, bsize, x, xd);
871 // Look into selecting tx_size here, based on prediction residual.
872 av1_block_yrd(x, &this_rdc, &args.skippable, bsize, mi->tx_size);
873 // TODO(kyslov@) Need to account for skippable
874 if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)]) {
875 av1_foreach_transformed_block_in_plane(xd, uv_bsize, AOM_PLANE_U,
876 av1_estimate_block_intra, &args);
877 }
878 if (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)]) {
879 av1_foreach_transformed_block_in_plane(xd, uv_bsize, AOM_PLANE_V,
880 av1_estimate_block_intra, &args);
881 }
882
883 int mode_cost = 0;
884 if (av1_is_directional_mode(this_mode) && av1_use_angle_delta(bsize)) {
885 mode_cost +=
886 x->mode_costs.angle_delta_cost[this_mode - V_PRED]
887 [MAX_ANGLE_DELTA +
888 mi->angle_delta[PLANE_TYPE_Y]];
889 }
890 if (this_mode == DC_PRED && av1_filter_intra_allowed_bsize(cm, bsize)) {
891 mode_cost += x->mode_costs.filter_intra_cost[bsize][0];
892 }
893 this_rdc.rate += ref_cost_intra;
894 this_rdc.rate += intra_cost_penalty;
895 this_rdc.rate += mode_cost;
896 this_rdc.rdcost = RDCOST(x->rdmult, this_rdc.rate, this_rdc.dist);
897
898 if (is_screen_content && rt_sf->source_metrics_sb_nonrd) {
899 // For blocks with low spatial variance and color sad,
900 // favor the intra-modes, only on scene/slide change.
901 if (cpi->rc.high_source_sad && x->source_variance < 800 &&
902 (x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] ||
903 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)]))
904 this_rdc.rdcost = CALC_BIASED_RDCOST(this_rdc.rdcost);
905 // Otherwise bias against intra for blocks with zero
906 // motion and no color, on non-scene/slide changes.
907 else if (!cpi->rc.high_source_sad && x->source_variance > 0 &&
908 x->content_state_sb.source_sad_nonrd == kZeroSad &&
909 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_U)] == 0 &&
910 x->color_sensitivity[COLOR_SENS_IDX(AOM_PLANE_V)] == 0)
911 this_rdc.rdcost = (3 * this_rdc.rdcost) >> 1;
912 }
913
914 if (this_rdc.rdcost < best_rdc->rdcost) {
915 *best_rdc = this_rdc;
916 best_pickmode->best_mode = this_mode;
917 best_pickmode->best_tx_size = mi->tx_size;
918 best_pickmode->best_ref_frame = INTRA_FRAME;
919 best_pickmode->best_second_ref_frame = NONE;
920 best_pickmode->best_mode_skip_txfm = this_rdc.skip_txfm;
921 mi->uv_mode = this_mode;
922 mi->mv[0].as_int = INVALID_MV;
923 mi->mv[1].as_int = INVALID_MV;
924 if (!this_rdc.skip_txfm)
925 memset(ctx->blk_skip, 0,
926 sizeof(x->txfm_search_info.blk_skip[0]) * ctx->num_4x4_blk);
927 }
928 }
929 if (best_pickmode->best_ref_frame == INTRA_FRAME)
930 memset(ctx->blk_skip, 0,
931 sizeof(x->txfm_search_info.blk_skip[0]) * ctx->num_4x4_blk);
932 mi->tx_size = best_pickmode->best_tx_size;
933 }
934