1 /* 2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11 #include "./vpx_config.h" 12 #include "./vpx_dsp_rtcd.h" 13 14 #if CONFIG_VP9_HIGHBITDEPTH 15 #include "vpx_dsp/vpx_dsp_common.h" 16 #endif // CONFIG_VP9_HIGHBITDEPTH 17 #include "vpx_mem/vpx_mem.h" 18 #include "vpx_ports/mem.h" 19 #include "vpx_ports/vpx_once.h" 20 21 #include "vp9/common/vp9_reconintra.h" 22 #include "vp9/common/vp9_onyxc_int.h" 23 24 const TX_TYPE intra_mode_to_tx_type_lookup[INTRA_MODES] = { 25 DCT_DCT, // DC 26 ADST_DCT, // V 27 DCT_ADST, // H 28 DCT_DCT, // D45 29 ADST_ADST, // D135 30 ADST_DCT, // D117 31 DCT_ADST, // D153 32 DCT_ADST, // D207 33 ADST_DCT, // D63 34 ADST_ADST, // TM 35 }; 36 37 enum { 38 NEED_LEFT = 1 << 1, 39 NEED_ABOVE = 1 << 2, 40 NEED_ABOVERIGHT = 1 << 3, 41 }; 42 43 static const uint8_t extend_modes[INTRA_MODES] = { 44 NEED_ABOVE | NEED_LEFT, // DC 45 NEED_ABOVE, // V 46 NEED_LEFT, // H 47 NEED_ABOVERIGHT, // D45 48 NEED_LEFT | NEED_ABOVE, // D135 49 NEED_LEFT | NEED_ABOVE, // D117 50 NEED_LEFT | NEED_ABOVE, // D153 51 NEED_LEFT, // D207 52 NEED_ABOVERIGHT, // D63 53 NEED_LEFT | NEED_ABOVE, // TM 54 }; 55 56 typedef void (*intra_pred_fn)(uint8_t *dst, ptrdiff_t stride, 57 const uint8_t *above, const uint8_t *left); 58 59 static intra_pred_fn pred[INTRA_MODES][TX_SIZES]; 60 static intra_pred_fn dc_pred[2][2][TX_SIZES]; 61 62 #if CONFIG_VP9_HIGHBITDEPTH 63 typedef void (*intra_high_pred_fn)(uint16_t *dst, ptrdiff_t stride, 64 const uint16_t *above, const uint16_t *left, 65 int bd); 66 static intra_high_pred_fn pred_high[INTRA_MODES][4]; 67 static intra_high_pred_fn dc_pred_high[2][2][4]; 68 #endif // CONFIG_VP9_HIGHBITDEPTH 69 70 static void vp9_init_intra_predictors_internal(void) { 71 #define INIT_ALL_SIZES(p, type) \ 72 p[TX_4X4] = vpx_##type##_predictor_4x4; \ 73 p[TX_8X8] = vpx_##type##_predictor_8x8; \ 74 p[TX_16X16] = vpx_##type##_predictor_16x16; \ 75 p[TX_32X32] = vpx_##type##_predictor_32x32 76 77 INIT_ALL_SIZES(pred[V_PRED], v); 78 INIT_ALL_SIZES(pred[H_PRED], h); 79 INIT_ALL_SIZES(pred[D207_PRED], d207); 80 INIT_ALL_SIZES(pred[D45_PRED], d45); 81 INIT_ALL_SIZES(pred[D63_PRED], d63); 82 INIT_ALL_SIZES(pred[D117_PRED], d117); 83 INIT_ALL_SIZES(pred[D135_PRED], d135); 84 INIT_ALL_SIZES(pred[D153_PRED], d153); 85 INIT_ALL_SIZES(pred[TM_PRED], tm); 86 87 INIT_ALL_SIZES(dc_pred[0][0], dc_128); 88 INIT_ALL_SIZES(dc_pred[0][1], dc_top); 89 INIT_ALL_SIZES(dc_pred[1][0], dc_left); 90 INIT_ALL_SIZES(dc_pred[1][1], dc); 91 92 #if CONFIG_VP9_HIGHBITDEPTH 93 INIT_ALL_SIZES(pred_high[V_PRED], highbd_v); 94 INIT_ALL_SIZES(pred_high[H_PRED], highbd_h); 95 INIT_ALL_SIZES(pred_high[D207_PRED], highbd_d207); 96 INIT_ALL_SIZES(pred_high[D45_PRED], highbd_d45); 97 INIT_ALL_SIZES(pred_high[D63_PRED], highbd_d63); 98 INIT_ALL_SIZES(pred_high[D117_PRED], highbd_d117); 99 INIT_ALL_SIZES(pred_high[D135_PRED], highbd_d135); 100 INIT_ALL_SIZES(pred_high[D153_PRED], highbd_d153); 101 INIT_ALL_SIZES(pred_high[TM_PRED], highbd_tm); 102 103 INIT_ALL_SIZES(dc_pred_high[0][0], highbd_dc_128); 104 INIT_ALL_SIZES(dc_pred_high[0][1], highbd_dc_top); 105 INIT_ALL_SIZES(dc_pred_high[1][0], highbd_dc_left); 106 INIT_ALL_SIZES(dc_pred_high[1][1], highbd_dc); 107 #endif // CONFIG_VP9_HIGHBITDEPTH 108 109 #undef intra_pred_allsizes 110 } 111 112 #if CONFIG_VP9_HIGHBITDEPTH 113 static void build_intra_predictors_high(const MACROBLOCKD *xd, 114 const uint8_t *ref8, 115 int ref_stride, 116 uint8_t *dst8, 117 int dst_stride, 118 PREDICTION_MODE mode, 119 TX_SIZE tx_size, 120 int up_available, 121 int left_available, 122 int right_available, 123 int x, int y, 124 int plane, int bd) { 125 int i; 126 uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); 127 uint16_t *ref = CONVERT_TO_SHORTPTR(ref8); 128 DECLARE_ALIGNED(16, uint16_t, left_col[32]); 129 DECLARE_ALIGNED(16, uint16_t, above_data[64 + 16]); 130 uint16_t *above_row = above_data + 16; 131 const uint16_t *const_above_row = above_row; 132 const int bs = 4 << tx_size; 133 int frame_width, frame_height; 134 int x0, y0; 135 const struct macroblockd_plane *const pd = &xd->plane[plane]; 136 int base = 128 << (bd - 8); 137 // 127 127 127 .. 127 127 127 127 127 127 138 // 129 A B .. Y Z 139 // 129 C D .. W X 140 // 129 E F .. U V 141 // 129 G H .. S T T T T T 142 143 // Get current frame pointer, width and height. 144 if (plane == 0) { 145 frame_width = xd->cur_buf->y_width; 146 frame_height = xd->cur_buf->y_height; 147 } else { 148 frame_width = xd->cur_buf->uv_width; 149 frame_height = xd->cur_buf->uv_height; 150 } 151 152 // Get block position in current frame. 153 x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x; 154 y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y; 155 156 // left 157 if (left_available) { 158 if (xd->mb_to_bottom_edge < 0) { 159 /* slower path if the block needs border extension */ 160 if (y0 + bs <= frame_height) { 161 for (i = 0; i < bs; ++i) 162 left_col[i] = ref[i * ref_stride - 1]; 163 } else { 164 const int extend_bottom = frame_height - y0; 165 for (i = 0; i < extend_bottom; ++i) 166 left_col[i] = ref[i * ref_stride - 1]; 167 for (; i < bs; ++i) 168 left_col[i] = ref[(extend_bottom - 1) * ref_stride - 1]; 169 } 170 } else { 171 /* faster path if the block does not need extension */ 172 for (i = 0; i < bs; ++i) 173 left_col[i] = ref[i * ref_stride - 1]; 174 } 175 } else { 176 // TODO(Peter): this value should probably change for high bitdepth 177 vpx_memset16(left_col, base + 1, bs); 178 } 179 180 // TODO(hkuang) do not extend 2*bs pixels for all modes. 181 // above 182 if (up_available) { 183 const uint16_t *above_ref = ref - ref_stride; 184 if (xd->mb_to_right_edge < 0) { 185 /* slower path if the block needs border extension */ 186 if (x0 + 2 * bs <= frame_width) { 187 if (right_available && bs == 4) { 188 memcpy(above_row, above_ref, 2 * bs * sizeof(above_row[0])); 189 } else { 190 memcpy(above_row, above_ref, bs * sizeof(above_row[0])); 191 vpx_memset16(above_row + bs, above_row[bs - 1], bs); 192 } 193 } else if (x0 + bs <= frame_width) { 194 const int r = frame_width - x0; 195 if (right_available && bs == 4) { 196 memcpy(above_row, above_ref, r * sizeof(above_row[0])); 197 vpx_memset16(above_row + r, above_row[r - 1], 198 x0 + 2 * bs - frame_width); 199 } else { 200 memcpy(above_row, above_ref, bs * sizeof(above_row[0])); 201 vpx_memset16(above_row + bs, above_row[bs - 1], bs); 202 } 203 } else if (x0 <= frame_width) { 204 const int r = frame_width - x0; 205 memcpy(above_row, above_ref, r * sizeof(above_row[0])); 206 vpx_memset16(above_row + r, above_row[r - 1], 207 x0 + 2 * bs - frame_width); 208 } 209 // TODO(Peter) this value should probably change for high bitdepth 210 above_row[-1] = left_available ? above_ref[-1] : (base+1); 211 } else { 212 /* faster path if the block does not need extension */ 213 if (bs == 4 && right_available && left_available) { 214 const_above_row = above_ref; 215 } else { 216 memcpy(above_row, above_ref, bs * sizeof(above_row[0])); 217 if (bs == 4 && right_available) 218 memcpy(above_row + bs, above_ref + bs, bs * sizeof(above_row[0])); 219 else 220 vpx_memset16(above_row + bs, above_row[bs - 1], bs); 221 // TODO(Peter): this value should probably change for high bitdepth 222 above_row[-1] = left_available ? above_ref[-1] : (base+1); 223 } 224 } 225 } else { 226 vpx_memset16(above_row, base - 1, bs * 2); 227 // TODO(Peter): this value should probably change for high bitdepth 228 above_row[-1] = base - 1; 229 } 230 231 // predict 232 if (mode == DC_PRED) { 233 dc_pred_high[left_available][up_available][tx_size](dst, dst_stride, 234 const_above_row, 235 left_col, xd->bd); 236 } else { 237 pred_high[mode][tx_size](dst, dst_stride, const_above_row, left_col, 238 xd->bd); 239 } 240 } 241 #endif // CONFIG_VP9_HIGHBITDEPTH 242 243 static void build_intra_predictors(const MACROBLOCKD *xd, const uint8_t *ref, 244 int ref_stride, uint8_t *dst, int dst_stride, 245 PREDICTION_MODE mode, TX_SIZE tx_size, 246 int up_available, int left_available, 247 int right_available, int x, int y, 248 int plane) { 249 int i; 250 DECLARE_ALIGNED(16, uint8_t, left_col[32]); 251 DECLARE_ALIGNED(16, uint8_t, above_data[64 + 16]); 252 uint8_t *above_row = above_data + 16; 253 const uint8_t *const_above_row = above_row; 254 const int bs = 4 << tx_size; 255 int frame_width, frame_height; 256 int x0, y0; 257 const struct macroblockd_plane *const pd = &xd->plane[plane]; 258 259 // 127 127 127 .. 127 127 127 127 127 127 260 // 129 A B .. Y Z 261 // 129 C D .. W X 262 // 129 E F .. U V 263 // 129 G H .. S T T T T T 264 // .. 265 266 // Get current frame pointer, width and height. 267 if (plane == 0) { 268 frame_width = xd->cur_buf->y_width; 269 frame_height = xd->cur_buf->y_height; 270 } else { 271 frame_width = xd->cur_buf->uv_width; 272 frame_height = xd->cur_buf->uv_height; 273 } 274 275 // Get block position in current frame. 276 x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x; 277 y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y; 278 279 // NEED_LEFT 280 if (extend_modes[mode] & NEED_LEFT) { 281 if (left_available) { 282 if (xd->mb_to_bottom_edge < 0) { 283 /* slower path if the block needs border extension */ 284 if (y0 + bs <= frame_height) { 285 for (i = 0; i < bs; ++i) 286 left_col[i] = ref[i * ref_stride - 1]; 287 } else { 288 const int extend_bottom = frame_height - y0; 289 for (i = 0; i < extend_bottom; ++i) 290 left_col[i] = ref[i * ref_stride - 1]; 291 for (; i < bs; ++i) 292 left_col[i] = ref[(extend_bottom - 1) * ref_stride - 1]; 293 } 294 } else { 295 /* faster path if the block does not need extension */ 296 for (i = 0; i < bs; ++i) 297 left_col[i] = ref[i * ref_stride - 1]; 298 } 299 } else { 300 memset(left_col, 129, bs); 301 } 302 } 303 304 // NEED_ABOVE 305 if (extend_modes[mode] & NEED_ABOVE) { 306 if (up_available) { 307 const uint8_t *above_ref = ref - ref_stride; 308 if (xd->mb_to_right_edge < 0) { 309 /* slower path if the block needs border extension */ 310 if (x0 + bs <= frame_width) { 311 memcpy(above_row, above_ref, bs); 312 } else if (x0 <= frame_width) { 313 const int r = frame_width - x0; 314 memcpy(above_row, above_ref, r); 315 memset(above_row + r, above_row[r - 1], x0 + bs - frame_width); 316 } 317 } else { 318 /* faster path if the block does not need extension */ 319 if (bs == 4 && right_available && left_available) { 320 const_above_row = above_ref; 321 } else { 322 memcpy(above_row, above_ref, bs); 323 } 324 } 325 above_row[-1] = left_available ? above_ref[-1] : 129; 326 } else { 327 memset(above_row, 127, bs); 328 above_row[-1] = 127; 329 } 330 } 331 332 // NEED_ABOVERIGHT 333 if (extend_modes[mode] & NEED_ABOVERIGHT) { 334 if (up_available) { 335 const uint8_t *above_ref = ref - ref_stride; 336 if (xd->mb_to_right_edge < 0) { 337 /* slower path if the block needs border extension */ 338 if (x0 + 2 * bs <= frame_width) { 339 if (right_available && bs == 4) { 340 memcpy(above_row, above_ref, 2 * bs); 341 } else { 342 memcpy(above_row, above_ref, bs); 343 memset(above_row + bs, above_row[bs - 1], bs); 344 } 345 } else if (x0 + bs <= frame_width) { 346 const int r = frame_width - x0; 347 if (right_available && bs == 4) { 348 memcpy(above_row, above_ref, r); 349 memset(above_row + r, above_row[r - 1], x0 + 2 * bs - frame_width); 350 } else { 351 memcpy(above_row, above_ref, bs); 352 memset(above_row + bs, above_row[bs - 1], bs); 353 } 354 } else if (x0 <= frame_width) { 355 const int r = frame_width - x0; 356 memcpy(above_row, above_ref, r); 357 memset(above_row + r, above_row[r - 1], x0 + 2 * bs - frame_width); 358 } 359 } else { 360 /* faster path if the block does not need extension */ 361 if (bs == 4 && right_available && left_available) { 362 const_above_row = above_ref; 363 } else { 364 memcpy(above_row, above_ref, bs); 365 if (bs == 4 && right_available) 366 memcpy(above_row + bs, above_ref + bs, bs); 367 else 368 memset(above_row + bs, above_row[bs - 1], bs); 369 } 370 } 371 above_row[-1] = left_available ? above_ref[-1] : 129; 372 } else { 373 memset(above_row, 127, bs * 2); 374 above_row[-1] = 127; 375 } 376 } 377 378 // predict 379 if (mode == DC_PRED) { 380 dc_pred[left_available][up_available][tx_size](dst, dst_stride, 381 const_above_row, left_col); 382 } else { 383 pred[mode][tx_size](dst, dst_stride, const_above_row, left_col); 384 } 385 } 386 387 void vp9_predict_intra_block(const MACROBLOCKD *xd, int bwl_in, 388 TX_SIZE tx_size, PREDICTION_MODE mode, 389 const uint8_t *ref, int ref_stride, 390 uint8_t *dst, int dst_stride, 391 int aoff, int loff, int plane) { 392 const int bw = (1 << bwl_in); 393 const int txw = (1 << tx_size); 394 const int have_top = loff || xd->up_available; 395 const int have_left = aoff || xd->left_available; 396 const int have_right = (aoff + txw) < bw; 397 const int x = aoff * 4; 398 const int y = loff * 4; 399 400 #if CONFIG_VP9_HIGHBITDEPTH 401 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) { 402 build_intra_predictors_high(xd, ref, ref_stride, dst, dst_stride, mode, 403 tx_size, have_top, have_left, have_right, 404 x, y, plane, xd->bd); 405 return; 406 } 407 #endif 408 build_intra_predictors(xd, ref, ref_stride, dst, dst_stride, mode, tx_size, 409 have_top, have_left, have_right, x, y, plane); 410 } 411 412 void vp9_init_intra_predictors(void) { 413 once(vp9_init_intra_predictors_internal); 414 } 415
