1 /* 2 * Copyright (c) 2015 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 <stdlib.h> 12 13 #include "./vpx_config.h" 14 #include "vpx_dsp/vpx_dsp_common.h" 15 #include "vpx_ports/mem.h" 16 17 static INLINE int8_t signed_char_clamp(int t) { 18 return (int8_t)clamp(t, -128, 127); 19 } 20 21 #if CONFIG_VP9_HIGHBITDEPTH 22 static INLINE int16_t signed_char_clamp_high(int t, int bd) { 23 switch (bd) { 24 case 10: 25 return (int16_t)clamp(t, -128*4, 128*4-1); 26 case 12: 27 return (int16_t)clamp(t, -128*16, 128*16-1); 28 case 8: 29 default: 30 return (int16_t)clamp(t, -128, 128-1); 31 } 32 } 33 #endif 34 35 // should we apply any filter at all: 11111111 yes, 00000000 no 36 static INLINE int8_t filter_mask(uint8_t limit, uint8_t blimit, 37 uint8_t p3, uint8_t p2, 38 uint8_t p1, uint8_t p0, 39 uint8_t q0, uint8_t q1, 40 uint8_t q2, uint8_t q3) { 41 int8_t mask = 0; 42 mask |= (abs(p3 - p2) > limit) * -1; 43 mask |= (abs(p2 - p1) > limit) * -1; 44 mask |= (abs(p1 - p0) > limit) * -1; 45 mask |= (abs(q1 - q0) > limit) * -1; 46 mask |= (abs(q2 - q1) > limit) * -1; 47 mask |= (abs(q3 - q2) > limit) * -1; 48 mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1; 49 return ~mask; 50 } 51 52 static INLINE int8_t flat_mask4(uint8_t thresh, 53 uint8_t p3, uint8_t p2, 54 uint8_t p1, uint8_t p0, 55 uint8_t q0, uint8_t q1, 56 uint8_t q2, uint8_t q3) { 57 int8_t mask = 0; 58 mask |= (abs(p1 - p0) > thresh) * -1; 59 mask |= (abs(q1 - q0) > thresh) * -1; 60 mask |= (abs(p2 - p0) > thresh) * -1; 61 mask |= (abs(q2 - q0) > thresh) * -1; 62 mask |= (abs(p3 - p0) > thresh) * -1; 63 mask |= (abs(q3 - q0) > thresh) * -1; 64 return ~mask; 65 } 66 67 static INLINE int8_t flat_mask5(uint8_t thresh, 68 uint8_t p4, uint8_t p3, 69 uint8_t p2, uint8_t p1, 70 uint8_t p0, uint8_t q0, 71 uint8_t q1, uint8_t q2, 72 uint8_t q3, uint8_t q4) { 73 int8_t mask = ~flat_mask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3); 74 mask |= (abs(p4 - p0) > thresh) * -1; 75 mask |= (abs(q4 - q0) > thresh) * -1; 76 return ~mask; 77 } 78 79 // is there high edge variance internal edge: 11111111 yes, 00000000 no 80 static INLINE int8_t hev_mask(uint8_t thresh, uint8_t p1, uint8_t p0, 81 uint8_t q0, uint8_t q1) { 82 int8_t hev = 0; 83 hev |= (abs(p1 - p0) > thresh) * -1; 84 hev |= (abs(q1 - q0) > thresh) * -1; 85 return hev; 86 } 87 88 static INLINE void filter4(int8_t mask, uint8_t thresh, uint8_t *op1, 89 uint8_t *op0, uint8_t *oq0, uint8_t *oq1) { 90 int8_t filter1, filter2; 91 92 const int8_t ps1 = (int8_t) *op1 ^ 0x80; 93 const int8_t ps0 = (int8_t) *op0 ^ 0x80; 94 const int8_t qs0 = (int8_t) *oq0 ^ 0x80; 95 const int8_t qs1 = (int8_t) *oq1 ^ 0x80; 96 const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1); 97 98 // add outer taps if we have high edge variance 99 int8_t filter = signed_char_clamp(ps1 - qs1) & hev; 100 101 // inner taps 102 filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask; 103 104 // save bottom 3 bits so that we round one side +4 and the other +3 105 // if it equals 4 we'll set to adjust by -1 to account for the fact 106 // we'd round 3 the other way 107 filter1 = signed_char_clamp(filter + 4) >> 3; 108 filter2 = signed_char_clamp(filter + 3) >> 3; 109 110 *oq0 = signed_char_clamp(qs0 - filter1) ^ 0x80; 111 *op0 = signed_char_clamp(ps0 + filter2) ^ 0x80; 112 113 // outer tap adjustments 114 filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; 115 116 *oq1 = signed_char_clamp(qs1 - filter) ^ 0x80; 117 *op1 = signed_char_clamp(ps1 + filter) ^ 0x80; 118 } 119 120 void vpx_lpf_horizontal_4_c(uint8_t *s, int p /* pitch */, 121 const uint8_t *blimit, const uint8_t *limit, 122 const uint8_t *thresh, int count) { 123 int i; 124 125 // loop filter designed to work using chars so that we can make maximum use 126 // of 8 bit simd instructions. 127 for (i = 0; i < 8 * count; ++i) { 128 const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p]; 129 const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p]; 130 const int8_t mask = filter_mask(*limit, *blimit, 131 p3, p2, p1, p0, q0, q1, q2, q3); 132 filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p); 133 ++s; 134 } 135 } 136 137 void vpx_lpf_horizontal_4_dual_c(uint8_t *s, int p, const uint8_t *blimit0, 138 const uint8_t *limit0, const uint8_t *thresh0, 139 const uint8_t *blimit1, const uint8_t *limit1, 140 const uint8_t *thresh1) { 141 vpx_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0, 1); 142 vpx_lpf_horizontal_4_c(s + 8, p, blimit1, limit1, thresh1, 1); 143 } 144 145 void vpx_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, 146 const uint8_t *limit, const uint8_t *thresh, 147 int count) { 148 int i; 149 150 // loop filter designed to work using chars so that we can make maximum use 151 // of 8 bit simd instructions. 152 for (i = 0; i < 8 * count; ++i) { 153 const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1]; 154 const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3]; 155 const int8_t mask = filter_mask(*limit, *blimit, 156 p3, p2, p1, p0, q0, q1, q2, q3); 157 filter4(mask, *thresh, s - 2, s - 1, s, s + 1); 158 s += pitch; 159 } 160 } 161 162 void vpx_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, 163 const uint8_t *limit0, const uint8_t *thresh0, 164 const uint8_t *blimit1, const uint8_t *limit1, 165 const uint8_t *thresh1) { 166 vpx_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0, 1); 167 vpx_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1, 168 thresh1, 1); 169 } 170 171 static INLINE void filter8(int8_t mask, uint8_t thresh, uint8_t flat, 172 uint8_t *op3, uint8_t *op2, 173 uint8_t *op1, uint8_t *op0, 174 uint8_t *oq0, uint8_t *oq1, 175 uint8_t *oq2, uint8_t *oq3) { 176 if (flat && mask) { 177 const uint8_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0; 178 const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3; 179 180 // 7-tap filter [1, 1, 1, 2, 1, 1, 1] 181 *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3); 182 *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3); 183 *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3); 184 *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3); 185 *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3); 186 *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3); 187 } else { 188 filter4(mask, thresh, op1, op0, oq0, oq1); 189 } 190 } 191 192 void vpx_lpf_horizontal_8_c(uint8_t *s, int p, const uint8_t *blimit, 193 const uint8_t *limit, const uint8_t *thresh, 194 int count) { 195 int i; 196 197 // loop filter designed to work using chars so that we can make maximum use 198 // of 8 bit simd instructions. 199 for (i = 0; i < 8 * count; ++i) { 200 const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p]; 201 const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p]; 202 203 const int8_t mask = filter_mask(*limit, *blimit, 204 p3, p2, p1, p0, q0, q1, q2, q3); 205 const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3); 206 filter8(mask, *thresh, flat, s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p, 207 s, s + 1 * p, s + 2 * p, s + 3 * p); 208 ++s; 209 } 210 } 211 212 void vpx_lpf_horizontal_8_dual_c(uint8_t *s, int p, const uint8_t *blimit0, 213 const uint8_t *limit0, const uint8_t *thresh0, 214 const uint8_t *blimit1, const uint8_t *limit1, 215 const uint8_t *thresh1) { 216 vpx_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0, 1); 217 vpx_lpf_horizontal_8_c(s + 8, p, blimit1, limit1, thresh1, 1); 218 } 219 220 void vpx_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, 221 const uint8_t *limit, const uint8_t *thresh, 222 int count) { 223 int i; 224 225 for (i = 0; i < 8 * count; ++i) { 226 const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1]; 227 const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3]; 228 const int8_t mask = filter_mask(*limit, *blimit, 229 p3, p2, p1, p0, q0, q1, q2, q3); 230 const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3); 231 filter8(mask, *thresh, flat, s - 4, s - 3, s - 2, s - 1, 232 s, s + 1, s + 2, s + 3); 233 s += pitch; 234 } 235 } 236 237 void vpx_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, 238 const uint8_t *limit0, const uint8_t *thresh0, 239 const uint8_t *blimit1, const uint8_t *limit1, 240 const uint8_t *thresh1) { 241 vpx_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0, 1); 242 vpx_lpf_vertical_8_c(s + 8 * pitch, pitch, blimit1, limit1, 243 thresh1, 1); 244 } 245 246 static INLINE void filter16(int8_t mask, uint8_t thresh, 247 uint8_t flat, uint8_t flat2, 248 uint8_t *op7, uint8_t *op6, 249 uint8_t *op5, uint8_t *op4, 250 uint8_t *op3, uint8_t *op2, 251 uint8_t *op1, uint8_t *op0, 252 uint8_t *oq0, uint8_t *oq1, 253 uint8_t *oq2, uint8_t *oq3, 254 uint8_t *oq4, uint8_t *oq5, 255 uint8_t *oq6, uint8_t *oq7) { 256 if (flat2 && flat && mask) { 257 const uint8_t p7 = *op7, p6 = *op6, p5 = *op5, p4 = *op4, 258 p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0; 259 260 const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3, 261 q4 = *oq4, q5 = *oq5, q6 = *oq6, q7 = *oq7; 262 263 // 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1] 264 *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 + 265 q0, 4); 266 *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 + 267 q0 + q1, 4); 268 *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 + 269 q0 + q1 + q2, 4); 270 *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 + 271 q0 + q1 + q2 + q3, 4); 272 *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 + 273 q0 + q1 + q2 + q3 + q4, 4); 274 *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 + 275 q0 + q1 + q2 + q3 + q4 + q5, 4); 276 *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 + 277 q0 + q1 + q2 + q3 + q4 + q5 + q6, 4); 278 *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 + 279 q0 * 2 + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4); 280 *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 + 281 q0 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4); 282 *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 + 283 q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 4); 284 *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 + 285 q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4); 286 *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 287 q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4); 288 *oq5 = ROUND_POWER_OF_TWO(p1 + p0 + 289 q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4); 290 *oq6 = ROUND_POWER_OF_TWO(p0 + 291 q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4); 292 } else { 293 filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3); 294 } 295 } 296 297 void vpx_lpf_horizontal_16_c(uint8_t *s, int p, const uint8_t *blimit, 298 const uint8_t *limit, const uint8_t *thresh, 299 int count) { 300 int i; 301 302 // loop filter designed to work using chars so that we can make maximum use 303 // of 8 bit simd instructions. 304 for (i = 0; i < 8 * count; ++i) { 305 const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p]; 306 const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p]; 307 const int8_t mask = filter_mask(*limit, *blimit, 308 p3, p2, p1, p0, q0, q1, q2, q3); 309 const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3); 310 const int8_t flat2 = flat_mask5(1, 311 s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0, 312 q0, s[4 * p], s[5 * p], s[6 * p], s[7 * p]); 313 314 filter16(mask, *thresh, flat, flat2, 315 s - 8 * p, s - 7 * p, s - 6 * p, s - 5 * p, 316 s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p, 317 s, s + 1 * p, s + 2 * p, s + 3 * p, 318 s + 4 * p, s + 5 * p, s + 6 * p, s + 7 * p); 319 ++s; 320 } 321 } 322 323 static void mb_lpf_vertical_edge_w(uint8_t *s, int p, 324 const uint8_t *blimit, 325 const uint8_t *limit, 326 const uint8_t *thresh, 327 int count) { 328 int i; 329 330 for (i = 0; i < count; ++i) { 331 const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1]; 332 const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3]; 333 const int8_t mask = filter_mask(*limit, *blimit, 334 p3, p2, p1, p0, q0, q1, q2, q3); 335 const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3); 336 const int8_t flat2 = flat_mask5(1, s[-8], s[-7], s[-6], s[-5], p0, 337 q0, s[4], s[5], s[6], s[7]); 338 339 filter16(mask, *thresh, flat, flat2, 340 s - 8, s - 7, s - 6, s - 5, s - 4, s - 3, s - 2, s - 1, 341 s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6, s + 7); 342 s += p; 343 } 344 } 345 346 void vpx_lpf_vertical_16_c(uint8_t *s, int p, const uint8_t *blimit, 347 const uint8_t *limit, const uint8_t *thresh) { 348 mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8); 349 } 350 351 void vpx_lpf_vertical_16_dual_c(uint8_t *s, int p, const uint8_t *blimit, 352 const uint8_t *limit, const uint8_t *thresh) { 353 mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16); 354 } 355 356 #if CONFIG_VP9_HIGHBITDEPTH 357 // Should we apply any filter at all: 11111111 yes, 00000000 no ? 358 static INLINE int8_t highbd_filter_mask(uint8_t limit, uint8_t blimit, 359 uint16_t p3, uint16_t p2, 360 uint16_t p1, uint16_t p0, 361 uint16_t q0, uint16_t q1, 362 uint16_t q2, uint16_t q3, int bd) { 363 int8_t mask = 0; 364 int16_t limit16 = (uint16_t)limit << (bd - 8); 365 int16_t blimit16 = (uint16_t)blimit << (bd - 8); 366 mask |= (abs(p3 - p2) > limit16) * -1; 367 mask |= (abs(p2 - p1) > limit16) * -1; 368 mask |= (abs(p1 - p0) > limit16) * -1; 369 mask |= (abs(q1 - q0) > limit16) * -1; 370 mask |= (abs(q2 - q1) > limit16) * -1; 371 mask |= (abs(q3 - q2) > limit16) * -1; 372 mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit16) * -1; 373 return ~mask; 374 } 375 376 static INLINE int8_t highbd_flat_mask4(uint8_t thresh, 377 uint16_t p3, uint16_t p2, 378 uint16_t p1, uint16_t p0, 379 uint16_t q0, uint16_t q1, 380 uint16_t q2, uint16_t q3, int bd) { 381 int8_t mask = 0; 382 int16_t thresh16 = (uint16_t)thresh << (bd - 8); 383 mask |= (abs(p1 - p0) > thresh16) * -1; 384 mask |= (abs(q1 - q0) > thresh16) * -1; 385 mask |= (abs(p2 - p0) > thresh16) * -1; 386 mask |= (abs(q2 - q0) > thresh16) * -1; 387 mask |= (abs(p3 - p0) > thresh16) * -1; 388 mask |= (abs(q3 - q0) > thresh16) * -1; 389 return ~mask; 390 } 391 392 static INLINE int8_t highbd_flat_mask5(uint8_t thresh, 393 uint16_t p4, uint16_t p3, 394 uint16_t p2, uint16_t p1, 395 uint16_t p0, uint16_t q0, 396 uint16_t q1, uint16_t q2, 397 uint16_t q3, uint16_t q4, int bd) { 398 int8_t mask = ~highbd_flat_mask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3, bd); 399 int16_t thresh16 = (uint16_t)thresh << (bd - 8); 400 mask |= (abs(p4 - p0) > thresh16) * -1; 401 mask |= (abs(q4 - q0) > thresh16) * -1; 402 return ~mask; 403 } 404 405 // Is there high edge variance internal edge: 406 // 11111111_11111111 yes, 00000000_00000000 no ? 407 static INLINE int16_t highbd_hev_mask(uint8_t thresh, uint16_t p1, uint16_t p0, 408 uint16_t q0, uint16_t q1, int bd) { 409 int16_t hev = 0; 410 int16_t thresh16 = (uint16_t)thresh << (bd - 8); 411 hev |= (abs(p1 - p0) > thresh16) * -1; 412 hev |= (abs(q1 - q0) > thresh16) * -1; 413 return hev; 414 } 415 416 static INLINE void highbd_filter4(int8_t mask, uint8_t thresh, uint16_t *op1, 417 uint16_t *op0, uint16_t *oq0, uint16_t *oq1, 418 int bd) { 419 int16_t filter1, filter2; 420 // ^0x80 equivalent to subtracting 0x80 from the values to turn them 421 // into -128 to +127 instead of 0 to 255. 422 int shift = bd - 8; 423 const int16_t ps1 = (int16_t)*op1 - (0x80 << shift); 424 const int16_t ps0 = (int16_t)*op0 - (0x80 << shift); 425 const int16_t qs0 = (int16_t)*oq0 - (0x80 << shift); 426 const int16_t qs1 = (int16_t)*oq1 - (0x80 << shift); 427 const uint16_t hev = highbd_hev_mask(thresh, *op1, *op0, *oq0, *oq1, bd); 428 429 // Add outer taps if we have high edge variance. 430 int16_t filter = signed_char_clamp_high(ps1 - qs1, bd) & hev; 431 432 // Inner taps. 433 filter = signed_char_clamp_high(filter + 3 * (qs0 - ps0), bd) & mask; 434 435 // Save bottom 3 bits so that we round one side +4 and the other +3 436 // if it equals 4 we'll set to adjust by -1 to account for the fact 437 // we'd round 3 the other way. 438 filter1 = signed_char_clamp_high(filter + 4, bd) >> 3; 439 filter2 = signed_char_clamp_high(filter + 3, bd) >> 3; 440 441 *oq0 = signed_char_clamp_high(qs0 - filter1, bd) + (0x80 << shift); 442 *op0 = signed_char_clamp_high(ps0 + filter2, bd) + (0x80 << shift); 443 444 // Outer tap adjustments. 445 filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; 446 447 *oq1 = signed_char_clamp_high(qs1 - filter, bd) + (0x80 << shift); 448 *op1 = signed_char_clamp_high(ps1 + filter, bd) + (0x80 << shift); 449 } 450 451 void vpx_highbd_lpf_horizontal_4_c(uint16_t *s, int p /* pitch */, 452 const uint8_t *blimit, const uint8_t *limit, 453 const uint8_t *thresh, int count, int bd) { 454 int i; 455 456 // loop filter designed to work using chars so that we can make maximum use 457 // of 8 bit simd instructions. 458 for (i = 0; i < 8 * count; ++i) { 459 const uint16_t p3 = s[-4 * p]; 460 const uint16_t p2 = s[-3 * p]; 461 const uint16_t p1 = s[-2 * p]; 462 const uint16_t p0 = s[-p]; 463 const uint16_t q0 = s[0 * p]; 464 const uint16_t q1 = s[1 * p]; 465 const uint16_t q2 = s[2 * p]; 466 const uint16_t q3 = s[3 * p]; 467 const int8_t mask = highbd_filter_mask(*limit, *blimit, 468 p3, p2, p1, p0, q0, q1, q2, q3, bd); 469 highbd_filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p, bd); 470 ++s; 471 } 472 } 473 474 void vpx_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int p, 475 const uint8_t *blimit0, 476 const uint8_t *limit0, 477 const uint8_t *thresh0, 478 const uint8_t *blimit1, 479 const uint8_t *limit1, 480 const uint8_t *thresh1, 481 int bd) { 482 vpx_highbd_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0, 1, bd); 483 vpx_highbd_lpf_horizontal_4_c(s + 8, p, blimit1, limit1, thresh1, 1, bd); 484 } 485 486 void vpx_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit, 487 const uint8_t *limit, const uint8_t *thresh, 488 int count, int bd) { 489 int i; 490 491 // loop filter designed to work using chars so that we can make maximum use 492 // of 8 bit simd instructions. 493 for (i = 0; i < 8 * count; ++i) { 494 const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1]; 495 const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3]; 496 const int8_t mask = highbd_filter_mask(*limit, *blimit, 497 p3, p2, p1, p0, q0, q1, q2, q3, bd); 498 highbd_filter4(mask, *thresh, s - 2, s - 1, s, s + 1, bd); 499 s += pitch; 500 } 501 } 502 503 void vpx_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch, 504 const uint8_t *blimit0, 505 const uint8_t *limit0, 506 const uint8_t *thresh0, 507 const uint8_t *blimit1, 508 const uint8_t *limit1, 509 const uint8_t *thresh1, 510 int bd) { 511 vpx_highbd_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0, 1, bd); 512 vpx_highbd_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1, 513 thresh1, 1, bd); 514 } 515 516 static INLINE void highbd_filter8(int8_t mask, uint8_t thresh, uint8_t flat, 517 uint16_t *op3, uint16_t *op2, 518 uint16_t *op1, uint16_t *op0, 519 uint16_t *oq0, uint16_t *oq1, 520 uint16_t *oq2, uint16_t *oq3, int bd) { 521 if (flat && mask) { 522 const uint16_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0; 523 const uint16_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3; 524 525 // 7-tap filter [1, 1, 1, 2, 1, 1, 1] 526 *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3); 527 *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3); 528 *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3); 529 *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3); 530 *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3); 531 *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3); 532 } else { 533 highbd_filter4(mask, thresh, op1, op0, oq0, oq1, bd); 534 } 535 } 536 537 void vpx_highbd_lpf_horizontal_8_c(uint16_t *s, int p, const uint8_t *blimit, 538 const uint8_t *limit, const uint8_t *thresh, 539 int count, int bd) { 540 int i; 541 542 // loop filter designed to work using chars so that we can make maximum use 543 // of 8 bit simd instructions. 544 for (i = 0; i < 8 * count; ++i) { 545 const uint16_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p]; 546 const uint16_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p]; 547 548 const int8_t mask = highbd_filter_mask(*limit, *blimit, 549 p3, p2, p1, p0, q0, q1, q2, q3, bd); 550 const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, 551 bd); 552 highbd_filter8(mask, *thresh, flat, 553 s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p, 554 s, s + 1 * p, s + 2 * p, s + 3 * p, bd); 555 ++s; 556 } 557 } 558 559 void vpx_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int p, 560 const uint8_t *blimit0, 561 const uint8_t *limit0, 562 const uint8_t *thresh0, 563 const uint8_t *blimit1, 564 const uint8_t *limit1, 565 const uint8_t *thresh1, 566 int bd) { 567 vpx_highbd_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0, 1, bd); 568 vpx_highbd_lpf_horizontal_8_c(s + 8, p, blimit1, limit1, thresh1, 1, bd); 569 } 570 571 void vpx_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit, 572 const uint8_t *limit, const uint8_t *thresh, 573 int count, int bd) { 574 int i; 575 576 for (i = 0; i < 8 * count; ++i) { 577 const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1]; 578 const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3]; 579 const int8_t mask = highbd_filter_mask(*limit, *blimit, 580 p3, p2, p1, p0, q0, q1, q2, q3, bd); 581 const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, 582 bd); 583 highbd_filter8(mask, *thresh, flat, 584 s - 4, s - 3, s - 2, s - 1, 585 s, s + 1, s + 2, s + 3, 586 bd); 587 s += pitch; 588 } 589 } 590 591 void vpx_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch, 592 const uint8_t *blimit0, 593 const uint8_t *limit0, 594 const uint8_t *thresh0, 595 const uint8_t *blimit1, 596 const uint8_t *limit1, 597 const uint8_t *thresh1, 598 int bd) { 599 vpx_highbd_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0, 1, bd); 600 vpx_highbd_lpf_vertical_8_c(s + 8 * pitch, pitch, blimit1, limit1, 601 thresh1, 1, bd); 602 } 603 604 static INLINE void highbd_filter16(int8_t mask, uint8_t thresh, 605 uint8_t flat, uint8_t flat2, 606 uint16_t *op7, uint16_t *op6, 607 uint16_t *op5, uint16_t *op4, 608 uint16_t *op3, uint16_t *op2, 609 uint16_t *op1, uint16_t *op0, 610 uint16_t *oq0, uint16_t *oq1, 611 uint16_t *oq2, uint16_t *oq3, 612 uint16_t *oq4, uint16_t *oq5, 613 uint16_t *oq6, uint16_t *oq7, int bd) { 614 if (flat2 && flat && mask) { 615 const uint16_t p7 = *op7; 616 const uint16_t p6 = *op6; 617 const uint16_t p5 = *op5; 618 const uint16_t p4 = *op4; 619 const uint16_t p3 = *op3; 620 const uint16_t p2 = *op2; 621 const uint16_t p1 = *op1; 622 const uint16_t p0 = *op0; 623 const uint16_t q0 = *oq0; 624 const uint16_t q1 = *oq1; 625 const uint16_t q2 = *oq2; 626 const uint16_t q3 = *oq3; 627 const uint16_t q4 = *oq4; 628 const uint16_t q5 = *oq5; 629 const uint16_t q6 = *oq6; 630 const uint16_t q7 = *oq7; 631 632 // 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1] 633 *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 + 634 q0, 4); 635 *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 + 636 q0 + q1, 4); 637 *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 + 638 q0 + q1 + q2, 4); 639 *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 + 640 q0 + q1 + q2 + q3, 4); 641 *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 + 642 q0 + q1 + q2 + q3 + q4, 4); 643 *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 + 644 q0 + q1 + q2 + q3 + q4 + q5, 4); 645 *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 + 646 q0 + q1 + q2 + q3 + q4 + q5 + q6, 4); 647 *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 + 648 q0 * 2 + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4); 649 *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 + 650 q0 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4); 651 *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 + 652 q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 4); 653 *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 + 654 q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4); 655 *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 656 q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4); 657 *oq5 = ROUND_POWER_OF_TWO(p1 + p0 + 658 q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4); 659 *oq6 = ROUND_POWER_OF_TWO(p0 + 660 q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4); 661 } else { 662 highbd_filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3, 663 bd); 664 } 665 } 666 667 void vpx_highbd_lpf_horizontal_16_c(uint16_t *s, int p, const uint8_t *blimit, 668 const uint8_t *limit, const uint8_t *thresh, 669 int count, int bd) { 670 int i; 671 672 // loop filter designed to work using chars so that we can make maximum use 673 // of 8 bit simd instructions. 674 for (i = 0; i < 8 * count; ++i) { 675 const uint16_t p3 = s[-4 * p]; 676 const uint16_t p2 = s[-3 * p]; 677 const uint16_t p1 = s[-2 * p]; 678 const uint16_t p0 = s[-p]; 679 const uint16_t q0 = s[0 * p]; 680 const uint16_t q1 = s[1 * p]; 681 const uint16_t q2 = s[2 * p]; 682 const uint16_t q3 = s[3 * p]; 683 const int8_t mask = highbd_filter_mask(*limit, *blimit, 684 p3, p2, p1, p0, q0, q1, q2, q3, bd); 685 const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, 686 bd); 687 const int8_t flat2 = highbd_flat_mask5( 688 1, s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0, 689 q0, s[4 * p], s[5 * p], s[6 * p], s[7 * p], bd); 690 691 highbd_filter16(mask, *thresh, flat, flat2, 692 s - 8 * p, s - 7 * p, s - 6 * p, s - 5 * p, 693 s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p, 694 s, s + 1 * p, s + 2 * p, s + 3 * p, 695 s + 4 * p, s + 5 * p, s + 6 * p, s + 7 * p, 696 bd); 697 ++s; 698 } 699 } 700 701 static void highbd_mb_lpf_vertical_edge_w(uint16_t *s, int p, 702 const uint8_t *blimit, 703 const uint8_t *limit, 704 const uint8_t *thresh, 705 int count, int bd) { 706 int i; 707 708 for (i = 0; i < count; ++i) { 709 const uint16_t p3 = s[-4]; 710 const uint16_t p2 = s[-3]; 711 const uint16_t p1 = s[-2]; 712 const uint16_t p0 = s[-1]; 713 const uint16_t q0 = s[0]; 714 const uint16_t q1 = s[1]; 715 const uint16_t q2 = s[2]; 716 const uint16_t q3 = s[3]; 717 const int8_t mask = highbd_filter_mask(*limit, *blimit, 718 p3, p2, p1, p0, q0, q1, q2, q3, bd); 719 const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3, 720 bd); 721 const int8_t flat2 = highbd_flat_mask5(1, s[-8], s[-7], s[-6], s[-5], p0, 722 q0, s[4], s[5], s[6], s[7], bd); 723 724 highbd_filter16(mask, *thresh, flat, flat2, 725 s - 8, s - 7, s - 6, s - 5, s - 4, s - 3, s - 2, s - 1, 726 s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6, s + 7, 727 bd); 728 s += p; 729 } 730 } 731 732 void vpx_highbd_lpf_vertical_16_c(uint16_t *s, int p, const uint8_t *blimit, 733 const uint8_t *limit, const uint8_t *thresh, 734 int bd) { 735 highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8, bd); 736 } 737 738 void vpx_highbd_lpf_vertical_16_dual_c(uint16_t *s, int p, 739 const uint8_t *blimit, 740 const uint8_t *limit, 741 const uint8_t *thresh, 742 int bd) { 743 highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16, bd); 744 } 745 #endif // CONFIG_VP9_HIGHBITDEPTH 746