cdef_tmpl.c (13802B)
1 /* 2 * Copyright © 2018, VideoLAN and dav1d authors 3 * Copyright © 2018, Two Orioles, LLC 4 * All rights reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions are met: 8 * 9 * 1. Redistributions of source code must retain the above copyright notice, this 10 * list of conditions and the following disclaimer. 11 * 12 * 2. Redistributions in binary form must reproduce the above copyright notice, 13 * this list of conditions and the following disclaimer in the documentation 14 * and/or other materials provided with the distribution. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 18 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 19 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR 20 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES 21 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 22 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND 23 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 24 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS 25 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 26 */ 27 28 #include "config.h" 29 30 #include <stdlib.h> 31 32 #include "common/intops.h" 33 34 #include "src/cdef.h" 35 #include "src/tables.h" 36 37 static inline int constrain(const int diff, const int threshold, 38 const int shift) 39 { 40 const int adiff = abs(diff); 41 return apply_sign(imin(adiff, imax(0, threshold - (adiff >> shift))), diff); 42 } 43 44 static inline void fill(int16_t *tmp, const ptrdiff_t stride, 45 const int w, const int h) 46 { 47 /* Use a value that's a large positive number when interpreted as unsigned, 48 * and a large negative number when interpreted as signed. */ 49 for (int y = 0; y < h; y++) { 50 for (int x = 0; x < w; x++) 51 tmp[x] = INT16_MIN; 52 tmp += stride; 53 } 54 } 55 56 static void padding(int16_t *tmp, const ptrdiff_t tmp_stride, 57 const pixel *src, const ptrdiff_t src_stride, 58 const pixel (*left)[2], 59 const pixel *top, const pixel *bottom, 60 const int w, const int h, const enum CdefEdgeFlags edges) 61 { 62 // fill extended input buffer 63 int x_start = -2, x_end = w + 2, y_start = -2, y_end = h + 2; 64 if (!(edges & CDEF_HAVE_TOP)) { 65 fill(tmp - 2 - 2 * tmp_stride, tmp_stride, w + 4, 2); 66 y_start = 0; 67 } 68 if (!(edges & CDEF_HAVE_BOTTOM)) { 69 fill(tmp + h * tmp_stride - 2, tmp_stride, w + 4, 2); 70 y_end -= 2; 71 } 72 if (!(edges & CDEF_HAVE_LEFT)) { 73 fill(tmp + y_start * tmp_stride - 2, tmp_stride, 2, y_end - y_start); 74 x_start = 0; 75 } 76 if (!(edges & CDEF_HAVE_RIGHT)) { 77 fill(tmp + y_start * tmp_stride + w, tmp_stride, 2, y_end - y_start); 78 x_end -= 2; 79 } 80 81 for (int y = y_start; y < 0; y++) { 82 for (int x = x_start; x < x_end; x++) 83 tmp[x + y * tmp_stride] = top[x]; 84 top += PXSTRIDE(src_stride); 85 } 86 for (int y = 0; y < h; y++) 87 for (int x = x_start; x < 0; x++) 88 tmp[x + y * tmp_stride] = left[y][2 + x]; 89 for (int y = 0; y < h; y++) { 90 for (int x = (y < h) ? 0 : x_start; x < x_end; x++) 91 tmp[x] = src[x]; 92 src += PXSTRIDE(src_stride); 93 tmp += tmp_stride; 94 } 95 for (int y = h; y < y_end; y++) { 96 for (int x = x_start; x < x_end; x++) 97 tmp[x] = bottom[x]; 98 bottom += PXSTRIDE(src_stride); 99 tmp += tmp_stride; 100 } 101 102 } 103 104 static NOINLINE void 105 cdef_filter_block_c(pixel *dst, const ptrdiff_t dst_stride, 106 const pixel (*left)[2], 107 const pixel *const top, const pixel *const bottom, 108 const int pri_strength, const int sec_strength, 109 const int dir, const int damping, const int w, int h, 110 const enum CdefEdgeFlags edges HIGHBD_DECL_SUFFIX) 111 { 112 const ptrdiff_t tmp_stride = 12; 113 assert((w == 4 || w == 8) && (h == 4 || h == 8)); 114 int16_t tmp_buf[144]; // 12*12 is the maximum value of tmp_stride * (h + 4) 115 int16_t *tmp = tmp_buf + 2 * tmp_stride + 2; 116 117 padding(tmp, tmp_stride, dst, dst_stride, left, top, bottom, w, h, edges); 118 119 if (pri_strength) { 120 const int bitdepth_min_8 = bitdepth_from_max(bitdepth_max) - 8; 121 const int pri_tap = 4 - ((pri_strength >> bitdepth_min_8) & 1); 122 const int pri_shift = imax(0, damping - ulog2(pri_strength)); 123 if (sec_strength) { 124 const int sec_shift = damping - ulog2(sec_strength); 125 do { 126 for (int x = 0; x < w; x++) { 127 const int px = dst[x]; 128 int sum = 0; 129 int max = px, min = px; 130 int pri_tap_k = pri_tap; 131 for (int k = 0; k < 2; k++) { 132 const int off1 = dav1d_cdef_directions[dir + 2][k]; // dir 133 const int p0 = tmp[x + off1]; 134 const int p1 = tmp[x - off1]; 135 sum += pri_tap_k * constrain(p0 - px, pri_strength, pri_shift); 136 sum += pri_tap_k * constrain(p1 - px, pri_strength, pri_shift); 137 // if pri_tap_k == 4 then it becomes 2 else it remains 3 138 pri_tap_k = (pri_tap_k & 3) | 2; 139 min = umin(p0, min); 140 max = imax(p0, max); 141 min = umin(p1, min); 142 max = imax(p1, max); 143 const int off2 = dav1d_cdef_directions[dir + 4][k]; // dir + 2 144 const int off3 = dav1d_cdef_directions[dir + 0][k]; // dir - 2 145 const int s0 = tmp[x + off2]; 146 const int s1 = tmp[x - off2]; 147 const int s2 = tmp[x + off3]; 148 const int s3 = tmp[x - off3]; 149 // sec_tap starts at 2 and becomes 1 150 const int sec_tap = 2 - k; 151 sum += sec_tap * constrain(s0 - px, sec_strength, sec_shift); 152 sum += sec_tap * constrain(s1 - px, sec_strength, sec_shift); 153 sum += sec_tap * constrain(s2 - px, sec_strength, sec_shift); 154 sum += sec_tap * constrain(s3 - px, sec_strength, sec_shift); 155 min = umin(s0, min); 156 max = imax(s0, max); 157 min = umin(s1, min); 158 max = imax(s1, max); 159 min = umin(s2, min); 160 max = imax(s2, max); 161 min = umin(s3, min); 162 max = imax(s3, max); 163 } 164 dst[x] = iclip(px + ((sum - (sum < 0) + 8) >> 4), min, max); 165 } 166 dst += PXSTRIDE(dst_stride); 167 tmp += tmp_stride; 168 } while (--h); 169 } else { // pri_strength only 170 do { 171 for (int x = 0; x < w; x++) { 172 const int px = dst[x]; 173 int sum = 0; 174 int pri_tap_k = pri_tap; 175 for (int k = 0; k < 2; k++) { 176 const int off = dav1d_cdef_directions[dir + 2][k]; // dir 177 const int p0 = tmp[x + off]; 178 const int p1 = tmp[x - off]; 179 sum += pri_tap_k * constrain(p0 - px, pri_strength, pri_shift); 180 sum += pri_tap_k * constrain(p1 - px, pri_strength, pri_shift); 181 pri_tap_k = (pri_tap_k & 3) | 2; 182 } 183 dst[x] = px + ((sum - (sum < 0) + 8) >> 4); 184 } 185 dst += PXSTRIDE(dst_stride); 186 tmp += tmp_stride; 187 } while (--h); 188 } 189 } else { // sec_strength only 190 assert(sec_strength); 191 const int sec_shift = damping - ulog2(sec_strength); 192 do { 193 for (int x = 0; x < w; x++) { 194 const int px = dst[x]; 195 int sum = 0; 196 for (int k = 0; k < 2; k++) { 197 const int off1 = dav1d_cdef_directions[dir + 4][k]; // dir + 2 198 const int off2 = dav1d_cdef_directions[dir + 0][k]; // dir - 2 199 const int s0 = tmp[x + off1]; 200 const int s1 = tmp[x - off1]; 201 const int s2 = tmp[x + off2]; 202 const int s3 = tmp[x - off2]; 203 const int sec_tap = 2 - k; 204 sum += sec_tap * constrain(s0 - px, sec_strength, sec_shift); 205 sum += sec_tap * constrain(s1 - px, sec_strength, sec_shift); 206 sum += sec_tap * constrain(s2 - px, sec_strength, sec_shift); 207 sum += sec_tap * constrain(s3 - px, sec_strength, sec_shift); 208 } 209 dst[x] = px + ((sum - (sum < 0) + 8) >> 4); 210 } 211 dst += PXSTRIDE(dst_stride); 212 tmp += tmp_stride; 213 } while (--h); 214 } 215 } 216 217 #define cdef_fn(w, h) \ 218 static void cdef_filter_block_##w##x##h##_c(pixel *const dst, \ 219 const ptrdiff_t stride, \ 220 const pixel (*left)[2], \ 221 const pixel *const top, \ 222 const pixel *const bottom, \ 223 const int pri_strength, \ 224 const int sec_strength, \ 225 const int dir, \ 226 const int damping, \ 227 const enum CdefEdgeFlags edges \ 228 HIGHBD_DECL_SUFFIX) \ 229 { \ 230 cdef_filter_block_c(dst, stride, left, top, bottom, \ 231 pri_strength, sec_strength, dir, damping, w, h, edges HIGHBD_TAIL_SUFFIX); \ 232 } 233 234 cdef_fn(4, 4); 235 cdef_fn(4, 8); 236 cdef_fn(8, 8); 237 238 static int cdef_find_dir_c(const pixel *img, const ptrdiff_t stride, 239 unsigned *const var HIGHBD_DECL_SUFFIX) 240 { 241 const int bitdepth_min_8 = bitdepth_from_max(bitdepth_max) - 8; 242 int partial_sum_hv[2][8] = { { 0 } }; 243 int partial_sum_diag[2][15] = { { 0 } }; 244 int partial_sum_alt[4][11] = { { 0 } }; 245 246 for (int y = 0; y < 8; y++) { 247 for (int x = 0; x < 8; x++) { 248 const int px = (img[x] >> bitdepth_min_8) - 128; 249 250 partial_sum_diag[0][ y + x ] += px; 251 partial_sum_alt [0][ y + (x >> 1)] += px; 252 partial_sum_hv [0][ y ] += px; 253 partial_sum_alt [1][3 + y - (x >> 1)] += px; 254 partial_sum_diag[1][7 + y - x ] += px; 255 partial_sum_alt [2][3 - (y >> 1) + x ] += px; 256 partial_sum_hv [1][ x ] += px; 257 partial_sum_alt [3][ (y >> 1) + x ] += px; 258 } 259 img += PXSTRIDE(stride); 260 } 261 262 unsigned cost[8] = { 0 }; 263 for (int n = 0; n < 8; n++) { 264 cost[2] += partial_sum_hv[0][n] * partial_sum_hv[0][n]; 265 cost[6] += partial_sum_hv[1][n] * partial_sum_hv[1][n]; 266 } 267 cost[2] *= 105; 268 cost[6] *= 105; 269 270 static const uint16_t div_table[7] = { 840, 420, 280, 210, 168, 140, 120 }; 271 for (int n = 0; n < 7; n++) { 272 const int d = div_table[n]; 273 cost[0] += (partial_sum_diag[0][n] * partial_sum_diag[0][n] + 274 partial_sum_diag[0][14 - n] * partial_sum_diag[0][14 - n]) * d; 275 cost[4] += (partial_sum_diag[1][n] * partial_sum_diag[1][n] + 276 partial_sum_diag[1][14 - n] * partial_sum_diag[1][14 - n]) * d; 277 } 278 cost[0] += partial_sum_diag[0][7] * partial_sum_diag[0][7] * 105; 279 cost[4] += partial_sum_diag[1][7] * partial_sum_diag[1][7] * 105; 280 281 for (int n = 0; n < 4; n++) { 282 unsigned *const cost_ptr = &cost[n * 2 + 1]; 283 for (int m = 0; m < 5; m++) 284 *cost_ptr += partial_sum_alt[n][3 + m] * partial_sum_alt[n][3 + m]; 285 *cost_ptr *= 105; 286 for (int m = 0; m < 3; m++) { 287 const int d = div_table[2 * m + 1]; 288 *cost_ptr += (partial_sum_alt[n][m] * partial_sum_alt[n][m] + 289 partial_sum_alt[n][10 - m] * partial_sum_alt[n][10 - m]) * d; 290 } 291 } 292 293 int best_dir = 0; 294 unsigned best_cost = cost[0]; 295 for (int n = 1; n < 8; n++) { 296 if (cost[n] > best_cost) { 297 best_cost = cost[n]; 298 best_dir = n; 299 } 300 } 301 302 *var = (best_cost - (cost[best_dir ^ 4])) >> 10; 303 return best_dir; 304 } 305 306 #if HAVE_ASM 307 #if ARCH_AARCH64 || ARCH_ARM 308 #include "src/arm/cdef.h" 309 #elif ARCH_PPC64LE 310 #include "src/ppc/cdef.h" 311 #elif ARCH_RISCV 312 #include "src/riscv/cdef.h" 313 #elif ARCH_X86 314 #include "src/x86/cdef.h" 315 #elif ARCH_LOONGARCH64 316 #include "src/loongarch/cdef.h" 317 #endif 318 #endif 319 320 COLD void bitfn(dav1d_cdef_dsp_init)(Dav1dCdefDSPContext *const c) { 321 c->dir = cdef_find_dir_c; 322 c->fb[0] = cdef_filter_block_8x8_c; 323 c->fb[1] = cdef_filter_block_4x8_c; 324 c->fb[2] = cdef_filter_block_4x4_c; 325 326 #if HAVE_ASM 327 #if ARCH_AARCH64 || ARCH_ARM 328 cdef_dsp_init_arm(c); 329 #elif ARCH_PPC64LE 330 cdef_dsp_init_ppc(c); 331 #elif ARCH_RISCV 332 cdef_dsp_init_riscv(c); 333 #elif ARCH_X86 334 cdef_dsp_init_x86(c); 335 #elif ARCH_LOONGARCH64 336 cdef_dsp_init_loongarch(c); 337 #endif 338 #endif 339 }