| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* | ||
| 2 | * VVC inter prediction DSP | ||
| 3 | * | ||
| 4 | * Copyright (C) 2022 Nuo Mi | ||
| 5 | * | ||
| 6 | * This file is part of FFmpeg. | ||
| 7 | * | ||
| 8 | * FFmpeg is free software; you can redistribute it and/or | ||
| 9 | * modify it under the terms of the GNU Lesser General Public | ||
| 10 | * License as published by the Free Software Foundation; either | ||
| 11 | * version 2.1 of the License, or (at your option) any later version. | ||
| 12 | * | ||
| 13 | * FFmpeg is distributed in the hope that it will be useful, | ||
| 14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
| 15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | ||
| 16 | * Lesser General Public License for more details. | ||
| 17 | * | ||
| 18 | * You should have received a copy of the GNU Lesser General Public | ||
| 19 | * License along with FFmpeg; if not, write to the Free Software | ||
| 20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | ||
| 21 | */ | ||
| 22 | |||
| 23 | #include "libavcodec/h26x/h2656_inter_template.c" | ||
| 24 | #include "libavutil/imgutils.h" | ||
| 25 | |||
| 26 | #define TMP_STRIDE EDGE_EMU_BUFFER_STRIDE | ||
| 27 | 65406 | static void av_always_inline FUNC(put_scaled)(uint8_t *_dst, const ptrdiff_t _dst_stride, | |
| 28 | const uint8_t *const _src, ptrdiff_t _src_stride, const int src_height, | ||
| 29 | const int _x, const int _y, const int dx, const int dy, | ||
| 30 | const int height, const int8_t *hf, const int8_t *vf, const int width, const int is_uni, const int is_chroma) | ||
| 31 | { | ||
| 32 | int16_t tmp_array[TMP_STRIDE * MAX_PB_SIZE]; | ||
| 33 | 65406 | int16_t *tmp = tmp_array; | |
| 34 | 65406 | pixel *dst = (pixel*)_dst; | |
| 35 | 65406 | int16_t *dst16 = (int16_t*)_dst; | |
| 36 | 65406 | const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel); | |
| 37 | 65406 | const ptrdiff_t src_stride = _src_stride / sizeof(pixel); | |
| 38 | 65406 | const int shift = FFMAX(2, 14 - BIT_DEPTH); | |
| 39 | 65406 | const int offset = 1 << (shift - 1); | |
| 40 |
2/2✓ Branch 0 taken 12262 times.
✓ Branch 1 taken 20441 times.
|
65406 | const int taps = is_chroma ? VVC_INTER_CHROMA_TAPS : VVC_INTER_LUMA_TAPS; |
| 41 |
2/2✓ Branch 0 taken 12262 times.
✓ Branch 1 taken 20441 times.
|
65406 | const int extra = is_chroma ? CHROMA_EXTRA : LUMA_EXTRA; |
| 42 |
2/2✓ Branch 0 taken 12262 times.
✓ Branch 1 taken 20441 times.
|
65406 | const int extra_before = is_chroma ? CHROMA_EXTRA_BEFORE : LUMA_EXTRA_BEFORE; |
| 43 | 65406 | const int shift1 = 6 - is_chroma; | |
| 44 | 65406 | const int shift2 = 4 + is_chroma; | |
| 45 | 65406 | const int x0 = SCALED_INT(_x); | |
| 46 | 65406 | const int y0 = SCALED_INT(_y); | |
| 47 | |||
| 48 |
2/2✓ Branch 0 taken 191000 times.
✓ Branch 1 taken 32703 times.
|
447406 | for (int i = 0; i < width; i++) { |
| 49 | 382000 | const int tx = _x + dx * i; | |
| 50 | 382000 | const int x = SCALED_INT(tx) - x0; | |
| 51 | 382000 | const int mx = av_zero_extend(tx >> shift1, shift2); | |
| 52 | 382000 | const int8_t *filter = hf + mx * taps; | |
| 53 | 382000 | const pixel *src = (pixel*)_src - extra_before * src_stride; | |
| 54 | |||
| 55 |
2/2✓ Branch 0 taken 2455364 times.
✓ Branch 1 taken 191000 times.
|
5292728 | for (int j = 0; j < src_height + extra; j++) { |
| 56 |
2/2✓ Branch 0 taken 718092 times.
✓ Branch 1 taken 1737272 times.
|
4910728 | tmp[j] = (is_chroma ? CHROMA_FILTER(src, 1) : LUMA_FILTER(src, 1)) >> (BIT_DEPTH - 8); |
| 57 | 4910728 | src += src_stride; | |
| 58 | } | ||
| 59 | 382000 | tmp += TMP_STRIDE; | |
| 60 | } | ||
| 61 | |||
| 62 |
2/2✓ Branch 0 taken 208640 times.
✓ Branch 1 taken 32703 times.
|
482686 | for (int i = 0; i < height; i++) { |
| 63 | 417280 | const int ty = _y + dy * i; | |
| 64 | 417280 | const int x = SCALED_INT(ty) - y0; | |
| 65 | 417280 | const int mx = av_zero_extend(ty >> shift1, shift2); | |
| 66 | 417280 | const int8_t *filter = vf + mx * taps; | |
| 67 | |||
| 68 | 417280 | tmp = tmp_array + extra_before; | |
| 69 |
2/2✓ Branch 0 taken 2656704 times.
✓ Branch 1 taken 208640 times.
|
5730688 | for (int j = 0; j < width; j++) { |
| 70 |
2/2✓ Branch 0 taken 885568 times.
✓ Branch 1 taken 1771136 times.
|
5313408 | const int val = (is_chroma ? CHROMA_FILTER(tmp, 1) : LUMA_FILTER(tmp, 1)) >> 6; |
| 71 |
2/2✓ Branch 0 taken 1820352 times.
✓ Branch 1 taken 836352 times.
|
5313408 | if (is_uni) |
| 72 | 3640704 | dst[j] = av_clip_pixel((val + offset) >> shift); | |
| 73 | else | ||
| 74 | 1672704 | dst16[j] = val; | |
| 75 | 5313408 | tmp += TMP_STRIDE; | |
| 76 | } | ||
| 77 |
2/2✓ Branch 0 taken 159280 times.
✓ Branch 1 taken 49360 times.
|
417280 | if (is_uni) |
| 78 | 318560 | dst += dst_stride; | |
| 79 | else | ||
| 80 | 98720 | dst16 += dst_stride; | |
| 81 | } | ||
| 82 | 65406 | } | |
| 83 | |||
| 84 | 7036 | static void FUNC(put_luma_scaled)(int16_t *_dst, | |
| 85 | const uint8_t *_src, ptrdiff_t _src_stride, const int src_height, | ||
| 86 | const int x, const int y, const int dx, const int dy, | ||
| 87 | const int height, const int8_t *hf, const int8_t *vf, const int width) | ||
| 88 | { | ||
| 89 | 7036 | FUNC(put_scaled)((uint8_t *)_dst, MAX_PB_SIZE * sizeof(pixel), _src, _src_stride, src_height, x, y, dx, dy, height, hf, vf, width, 0, 0); | |
| 90 | 7036 | } | |
| 91 | |||
| 92 | 4856 | static void FUNC(put_chroma_scaled)(int16_t *_dst, | |
| 93 | const uint8_t *_src, ptrdiff_t _src_stride, const int src_height, | ||
| 94 | const int x, const int y, const int dx, const int dy, | ||
| 95 | const int height, const int8_t *hf, const int8_t *vf, const int width) | ||
| 96 | { | ||
| 97 | 4856 | FUNC(put_scaled)((uint8_t *)_dst, MAX_PB_SIZE * sizeof(pixel), _src, _src_stride, src_height, x, y, dx, dy, height, hf, vf, width, 0, 1); | |
| 98 | 4856 | } | |
| 99 | |||
| 100 | 33846 | static void FUNC(put_uni_luma_scaled)(uint8_t *_dst, const ptrdiff_t _dst_stride, | |
| 101 | const uint8_t *_src, ptrdiff_t _src_stride, const int src_height, | ||
| 102 | const int x, const int y, const int dx, const int dy, | ||
| 103 | const int height, const int8_t *hf, const int8_t *vf, const int width) | ||
| 104 | { | ||
| 105 | 33846 | FUNC(put_scaled)(_dst, _dst_stride, _src, _src_stride, src_height, x, y, dx, dy, height, hf, vf, width, 1, 0); | |
| 106 | 33846 | } | |
| 107 | |||
| 108 | 19668 | static void FUNC(put_uni_chroma_scaled)(uint8_t *_dst, const ptrdiff_t _dst_stride, | |
| 109 | const uint8_t *_src, ptrdiff_t _src_stride, const int src_height, | ||
| 110 | const int x, const int y, const int dx, const int dy, | ||
| 111 | const int height, const int8_t *hf, const int8_t *vf, const int width) | ||
| 112 | { | ||
| 113 | 19668 | FUNC(put_scaled)(_dst, _dst_stride, _src, _src_stride, src_height, x, y, dx, dy, height, hf, vf, width, 1, 1); | |
| 114 | 19668 | } | |
| 115 | |||
| 116 | ✗ | static void av_always_inline FUNC(put_uni_w_scaled)(uint8_t *_dst, const ptrdiff_t _dst_stride, | |
| 117 | const uint8_t *const _src, ptrdiff_t _src_stride, const int src_height, | ||
| 118 | const int _x, const int _y, const int dx, const int dy, const int denom, const int wx, const int _ox, | ||
| 119 | const int height, const int8_t *hf, const int8_t *vf, const int width, const int is_chroma) | ||
| 120 | { | ||
| 121 | int16_t tmp_array[TMP_STRIDE * MAX_PB_SIZE]; | ||
| 122 | ✗ | int16_t *tmp = tmp_array; | |
| 123 | ✗ | pixel *dst = (pixel*)_dst; | |
| 124 | ✗ | const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel); | |
| 125 | ✗ | const ptrdiff_t src_stride = _src_stride / sizeof(pixel); | |
| 126 | ✗ | const int shift = FFMAX(2, 14 - BIT_DEPTH); | |
| 127 | ✗ | const int offset = 1 << (shift - 1); | |
| 128 | ✗ | const int ox = _ox * (1 << (BIT_DEPTH - 8)); | |
| 129 | ✗ | const int taps = is_chroma ? VVC_INTER_CHROMA_TAPS : VVC_INTER_LUMA_TAPS; | |
| 130 | ✗ | const int extra = is_chroma ? CHROMA_EXTRA : LUMA_EXTRA; | |
| 131 | ✗ | const int extra_before = is_chroma ? CHROMA_EXTRA_BEFORE : LUMA_EXTRA_BEFORE; | |
| 132 | ✗ | const int shift1 = 6 - is_chroma; | |
| 133 | ✗ | const int shift2 = 4 + is_chroma; | |
| 134 | ✗ | const int x0 = SCALED_INT(_x); | |
| 135 | ✗ | const int y0 = SCALED_INT(_y); | |
| 136 | |||
| 137 | ✗ | for (int i = 0; i < width; i++) { | |
| 138 | ✗ | const int tx = _x + dx * i; | |
| 139 | ✗ | const int x = SCALED_INT(tx) - x0; | |
| 140 | ✗ | const int mx = av_zero_extend(tx >> shift1, shift2); | |
| 141 | ✗ | const int8_t *filter = hf + mx * taps; | |
| 142 | ✗ | const pixel *src = (pixel*)_src - extra_before * src_stride; | |
| 143 | |||
| 144 | ✗ | for (int j = 0; j < src_height + extra; j++) { | |
| 145 | ✗ | tmp[j] = (is_chroma ? CHROMA_FILTER(src, 1) : LUMA_FILTER(src, 1)) >> (BIT_DEPTH - 8); | |
| 146 | ✗ | src += src_stride; | |
| 147 | } | ||
| 148 | ✗ | tmp += TMP_STRIDE; | |
| 149 | } | ||
| 150 | |||
| 151 | ✗ | for (int i = 0; i < height; i++) { | |
| 152 | ✗ | const int ty = _y + dy * i; | |
| 153 | ✗ | const int x = SCALED_INT(ty) - y0; | |
| 154 | ✗ | const int mx = av_zero_extend(ty >> shift1, shift2); | |
| 155 | ✗ | const int8_t *filter = vf + mx * taps; | |
| 156 | |||
| 157 | ✗ | tmp = tmp_array + extra_before; | |
| 158 | ✗ | for (int j = 0; j < width; j++) { | |
| 159 | ✗ | const int val = (is_chroma ? CHROMA_FILTER(tmp, 1) : LUMA_FILTER(tmp, 1)) >> 6; | |
| 160 | ✗ | dst[j] = av_clip_pixel(((wx * val + offset) >> shift) + ox); | |
| 161 | ✗ | tmp += TMP_STRIDE; | |
| 162 | } | ||
| 163 | ✗ | dst += dst_stride; | |
| 164 | } | ||
| 165 | ✗ | } | |
| 166 | |||
| 167 | ✗ | static void FUNC(put_uni_luma_w_scaled)(uint8_t *_dst, const ptrdiff_t _dst_stride, | |
| 168 | const uint8_t *_src, ptrdiff_t _src_stride, const int src_height, | ||
| 169 | const int x, const int y, const int dx, const int dy, const int denom, const int wx, const int ox, | ||
| 170 | const int height, const int8_t *hf, const int8_t *vf, const int width) | ||
| 171 | { | ||
| 172 | ✗ | FUNC(put_uni_w_scaled)(_dst, _dst_stride, _src, _src_stride, src_height, x, y, dx, dy, denom, wx, ox, height, hf, vf, width, 0); | |
| 173 | ✗ | } | |
| 174 | |||
| 175 | ✗ | static void FUNC(put_uni_chroma_w_scaled)(uint8_t *_dst, const ptrdiff_t _dst_stride, | |
| 176 | const uint8_t *_src, ptrdiff_t _src_stride, const int src_height, | ||
| 177 | const int x, const int y, const int dx, const int dy, const int denom, const int wx, const int ox, | ||
| 178 | const int height, const int8_t *hf, const int8_t *vf, const int width) | ||
| 179 | { | ||
| 180 | ✗ | FUNC(put_uni_w_scaled)(_dst, _dst_stride, _src, _src_stride, src_height, x, y, dx, dy, denom, wx, ox, height, hf, vf, width, 1); | |
| 181 | ✗ | } | |
| 182 | |||
| 183 | #undef TMP_STRIDE | ||
| 184 | |||
| 185 | 12865108 | static void FUNC(avg)(uint8_t *_dst, const ptrdiff_t _dst_stride, | |
| 186 | const int16_t *src0, const int16_t *src1, const int width, const int height) | ||
| 187 | { | ||
| 188 | 12865108 | pixel *dst = (pixel*)_dst; | |
| 189 | 12865108 | const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel); | |
| 190 | 12865108 | const int shift = FFMAX(3, 15 - BIT_DEPTH); | |
| 191 | 12865108 | const int offset = 1 << (shift - 1); | |
| 192 | |||
| 193 |
2/2✓ Branch 0 taken 58936738 times.
✓ Branch 1 taken 6432554 times.
|
130738584 | for (int y = 0; y < height; y++) { |
| 194 |
2/2✓ Branch 0 taken 755942052 times.
✓ Branch 1 taken 58936738 times.
|
1629757580 | for (int x = 0; x < width; x++) |
| 195 | 1511884104 | dst[x] = av_clip_pixel((src0[x] + src1[x] + offset) >> shift); | |
| 196 | 117873476 | src0 += MAX_PB_SIZE; | |
| 197 | 117873476 | src1 += MAX_PB_SIZE; | |
| 198 | 117873476 | dst += dst_stride; | |
| 199 | } | ||
| 200 | 12865108 | } | |
| 201 | |||
| 202 | 1336866 | static void FUNC(w_avg)(uint8_t *_dst, const ptrdiff_t _dst_stride, | |
| 203 | const int16_t *src0, const int16_t *src1, const int width, const int height, | ||
| 204 | const int denom, const int w0, const int w1, const int o0, const int o1) | ||
| 205 | { | ||
| 206 | 1336866 | pixel *dst = (pixel*)_dst; | |
| 207 | 1336866 | const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel); | |
| 208 | 1336866 | const int shift = denom + FFMAX(3, 15 - BIT_DEPTH); | |
| 209 | 1336866 | const int offset = ((o0 + o1) * (1 << (BIT_DEPTH - 8)) + 1) * (1 << (shift - 1)); | |
| 210 | |||
| 211 |
2/2✓ Branch 0 taken 4619374 times.
✓ Branch 1 taken 668433 times.
|
10575614 | for (int y = 0; y < height; y++) { |
| 212 |
2/2✓ Branch 0 taken 124831876 times.
✓ Branch 1 taken 4619374 times.
|
258902500 | for (int x = 0; x < width; x++) |
| 213 | 249663752 | dst[x] = av_clip_pixel((src0[x] * w0 + src1[x] * w1 + offset) >> shift); | |
| 214 | 9238748 | src0 += MAX_PB_SIZE; | |
| 215 | 9238748 | src1 += MAX_PB_SIZE; | |
| 216 | 9238748 | dst += dst_stride; | |
| 217 | } | ||
| 218 | 1336866 | } | |
| 219 | |||
| 220 | 87666 | static void FUNC(put_ciip)(uint8_t *_dst, const ptrdiff_t _dst_stride, | |
| 221 | const int width, const int height, | ||
| 222 | const uint8_t *_inter, const ptrdiff_t _inter_stride, const int intra_weight) | ||
| 223 | { | ||
| 224 | 87666 | pixel *dst = (pixel *)_dst; | |
| 225 | 87666 | pixel *inter = (pixel *)_inter; | |
| 226 | 87666 | const size_t dst_stride = _dst_stride / sizeof(pixel); | |
| 227 | 87666 | const size_t inter_stride = _inter_stride / sizeof(pixel); | |
| 228 | 87666 | const int inter_weight = 4 - intra_weight; | |
| 229 | |||
| 230 |
2/2✓ Branch 0 taken 533728 times.
✓ Branch 1 taken 43833 times.
|
1155122 | for (int y = 0; y < height; y++) { |
| 231 |
2/2✓ Branch 0 taken 7331040 times.
✓ Branch 1 taken 533728 times.
|
15729536 | for (int x = 0; x < width; x++) |
| 232 | 14662080 | dst[x] = (dst[x] * intra_weight + inter[x] * inter_weight + 2) >> 2; | |
| 233 | 1067456 | dst += dst_stride; | |
| 234 | 1067456 | inter += inter_stride; | |
| 235 | } | ||
| 236 | 87666 | } | |
| 237 | |||
| 238 | 156440 | static void FUNC(put_gpm)(uint8_t *_dst, ptrdiff_t dst_stride, | |
| 239 | const int width, const int height, | ||
| 240 | const int16_t *src0, const int16_t *src1, | ||
| 241 | const uint8_t *weights, const int step_x, const int step_y) | ||
| 242 | { | ||
| 243 | 156440 | const int shift = FFMAX(5, 17 - BIT_DEPTH); | |
| 244 | 156440 | const int offset = 1 << (shift - 1); | |
| 245 | 156440 | pixel *dst = (pixel *)_dst; | |
| 246 | |||
| 247 | 156440 | dst_stride /= sizeof(pixel); | |
| 248 |
2/2✓ Branch 0 taken 998800 times.
✓ Branch 1 taken 78220 times.
|
2154040 | for (int y = 0; y < height; y++) { |
| 249 |
2/2✓ Branch 0 taken 16058944 times.
✓ Branch 1 taken 998800 times.
|
34115488 | for (int x = 0; x < width; x++) { |
| 250 | 32117888 | const uint8_t w = weights[x * step_x]; | |
| 251 | 32117888 | dst[x] = av_clip_pixel((src0[x] * w + src1[x] * (8 - w) + offset) >> shift); | |
| 252 | } | ||
| 253 | 1997600 | dst += dst_stride; | |
| 254 | 1997600 | src0 += MAX_PB_SIZE; | |
| 255 | 1997600 | src1 += MAX_PB_SIZE; | |
| 256 | 1997600 | weights += step_y; | |
| 257 | } | ||
| 258 | 156440 | } | |
| 259 | |||
| 260 | //8.5.6.3.3 Luma integer sample fetching process, add one extra pad line | ||
| 261 | 8552872 | static void FUNC(bdof_fetch_samples)(int16_t *_dst, const uint8_t *_src, const ptrdiff_t _src_stride, | |
| 262 | const int x_frac, const int y_frac, const int width, const int height) | ||
| 263 | { | ||
| 264 | 8552872 | const int x_off = (x_frac >> 3) - 1; | |
| 265 | 8552872 | const int y_off = (y_frac >> 3) - 1; | |
| 266 | 8552872 | const ptrdiff_t src_stride = _src_stride / sizeof(pixel); | |
| 267 | 8552872 | const pixel *src = (pixel*)_src + (x_off) + y_off * src_stride; | |
| 268 | 8552872 | int16_t *dst = _dst - 1 - MAX_PB_SIZE; | |
| 269 | 8552872 | const int shift = 14 - BIT_DEPTH; | |
| 270 | 8552872 | const int bdof_width = width + 2 * BDOF_BORDER_EXT; | |
| 271 | |||
| 272 | // top | ||
| 273 |
2/2✓ Branch 0 taken 33924872 times.
✓ Branch 1 taken 4276436 times.
|
76402616 | for (int i = 0; i < bdof_width; i++) |
| 274 | 67849744 | dst[i] = src[i] << shift; | |
| 275 | |||
| 276 | 8552872 | dst += MAX_PB_SIZE; | |
| 277 | 8552872 | src += src_stride; | |
| 278 | |||
| 279 |
2/2✓ Branch 0 taken 25445120 times.
✓ Branch 1 taken 4276436 times.
|
59443112 | for (int i = 0; i < height; i++) { |
| 280 | 50890240 | dst[0] = src[0] << shift; | |
| 281 | 50890240 | dst[1 + width] = src[1 + width] << shift; | |
| 282 | 50890240 | dst += MAX_PB_SIZE; | |
| 283 | 50890240 | src += src_stride; | |
| 284 | } | ||
| 285 |
2/2✓ Branch 0 taken 33924872 times.
✓ Branch 1 taken 4276436 times.
|
76402616 | for (int i = 0; i < bdof_width; i++) |
| 286 | 67849744 | dst[i] = src[i] << shift; | |
| 287 | 8552872 | } | |
| 288 | |||
| 289 | //8.5.6.3.3 Luma integer sample fetching process | ||
| 290 | 7122672 | static void FUNC(fetch_samples)(int16_t *_dst, const uint8_t *_src, const ptrdiff_t _src_stride, const int x_frac, const int y_frac) | |
| 291 | { | ||
| 292 | 7122672 | FUNC(bdof_fetch_samples)(_dst, _src, _src_stride, x_frac, y_frac, AFFINE_MIN_BLOCK_SIZE, AFFINE_MIN_BLOCK_SIZE); | |
| 293 | 7122672 | } | |
| 294 | |||
| 295 | 8552980 | static void FUNC(prof_grad_filter)(int16_t *gradient_h, int16_t *gradient_v, const ptrdiff_t gradient_stride, | |
| 296 | const int16_t *_src, const ptrdiff_t src_stride, const int width, const int height) | ||
| 297 | { | ||
| 298 | 8552980 | const int shift = 6; | |
| 299 | 8552980 | const int16_t *src = _src; | |
| 300 | |||
| 301 |
2/2✓ Branch 0 taken 25445840 times.
✓ Branch 1 taken 4276490 times.
|
59444660 | for (int y = 0; y < height; y++) { |
| 302 | 50891680 | const int16_t *p = src; | |
| 303 |
2/2✓ Branch 0 taken 231147904 times.
✓ Branch 1 taken 25445840 times.
|
513187488 | for (int x = 0; x < width; x++) { |
| 304 | 462295808 | gradient_h[x] = (p[1] >> shift) - (p[-1] >> shift); | |
| 305 | 462295808 | gradient_v[x] = (p[src_stride] >> shift) - (p[-src_stride] >> shift); | |
| 306 | 462295808 | p++; | |
| 307 | } | ||
| 308 | 50891680 | gradient_h += gradient_stride; | |
| 309 | 50891680 | gradient_v += gradient_stride; | |
| 310 | 50891680 | src += src_stride; | |
| 311 | } | ||
| 312 | 8552980 | } | |
| 313 | |||
| 314 | 2820992 | static void FUNC(apply_prof)(int16_t *dst, const int16_t *src, const int16_t *diff_mv_x, const int16_t *diff_mv_y) | |
| 315 | { | ||
| 316 | 2820992 | const int limit = (1 << FFMAX(13, BIT_DEPTH + 1)); ///< dILimit | |
| 317 | |||
| 318 | int16_t gradient_h[AFFINE_MIN_BLOCK_SIZE * AFFINE_MIN_BLOCK_SIZE]; | ||
| 319 | int16_t gradient_v[AFFINE_MIN_BLOCK_SIZE * AFFINE_MIN_BLOCK_SIZE]; | ||
| 320 | 2820992 | FUNC(prof_grad_filter)(gradient_h, gradient_v, AFFINE_MIN_BLOCK_SIZE, src, MAX_PB_SIZE, AFFINE_MIN_BLOCK_SIZE, AFFINE_MIN_BLOCK_SIZE); | |
| 321 | |||
| 322 |
2/2✓ Branch 0 taken 5641984 times.
✓ Branch 1 taken 1410496 times.
|
14104960 | for (int y = 0; y < AFFINE_MIN_BLOCK_SIZE; y++) { |
| 323 |
2/2✓ Branch 0 taken 22567936 times.
✓ Branch 1 taken 5641984 times.
|
56419840 | for (int x = 0; x < AFFINE_MIN_BLOCK_SIZE; x++) { |
| 324 | 45135872 | const int o = y * AFFINE_MIN_BLOCK_SIZE + x; | |
| 325 | 45135872 | const int di = gradient_h[o] * diff_mv_x[o] + gradient_v[o] * diff_mv_y[o]; | |
| 326 | 45135872 | const int val = src[x] + av_clip(di, -limit, limit - 1); | |
| 327 | 45135872 | dst[x] = val; | |
| 328 | |||
| 329 | } | ||
| 330 | 11283968 | src += MAX_PB_SIZE; | |
| 331 | 11283968 | dst += MAX_PB_SIZE; | |
| 332 | } | ||
| 333 | 2820992 | } | |
| 334 | |||
| 335 | 4273016 | static void FUNC(apply_prof_uni)(uint8_t *_dst, const ptrdiff_t _dst_stride, const int16_t *src, const int16_t *diff_mv_x, const int16_t *diff_mv_y) | |
| 336 | { | ||
| 337 | 4273016 | const int limit = (1 << FFMAX(13, BIT_DEPTH + 1)); ///< dILimit | |
| 338 | 4273016 | pixel *dst = (pixel*)_dst; | |
| 339 | 4273016 | const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel); | |
| 340 | 4273016 | const int shift = 14 - BIT_DEPTH; | |
| 341 | #if BIT_DEPTH < 14 | ||
| 342 | 4273016 | const int offset = 1 << (shift - 1); | |
| 343 | #else | ||
| 344 | const int offset = 0; | ||
| 345 | #endif | ||
| 346 | int16_t gradient_h[AFFINE_MIN_BLOCK_SIZE * AFFINE_MIN_BLOCK_SIZE]; | ||
| 347 | int16_t gradient_v[AFFINE_MIN_BLOCK_SIZE * AFFINE_MIN_BLOCK_SIZE]; | ||
| 348 | |||
| 349 | 4273016 | FUNC(prof_grad_filter)(gradient_h, gradient_v, AFFINE_MIN_BLOCK_SIZE, src, MAX_PB_SIZE, AFFINE_MIN_BLOCK_SIZE, AFFINE_MIN_BLOCK_SIZE); | |
| 350 | |||
| 351 |
2/2✓ Branch 0 taken 8546032 times.
✓ Branch 1 taken 2136508 times.
|
21365080 | for (int y = 0; y < AFFINE_MIN_BLOCK_SIZE; y++) { |
| 352 |
2/2✓ Branch 0 taken 34184128 times.
✓ Branch 1 taken 8546032 times.
|
85460320 | for (int x = 0; x < AFFINE_MIN_BLOCK_SIZE; x++) { |
| 353 | 68368256 | const int o = y * AFFINE_MIN_BLOCK_SIZE + x; | |
| 354 | 68368256 | const int di = gradient_h[o] * diff_mv_x[o] + gradient_v[o] * diff_mv_y[o]; | |
| 355 | 68368256 | const int val = src[x] + av_clip(di, -limit, limit - 1); | |
| 356 | 68368256 | dst[x] = av_clip_pixel((val + offset) >> shift); | |
| 357 | |||
| 358 | } | ||
| 359 | 17092064 | src += MAX_PB_SIZE; | |
| 360 | 17092064 | dst += dst_stride; | |
| 361 | } | ||
| 362 | 4273016 | } | |
| 363 | |||
| 364 | 28664 | static void FUNC(apply_prof_uni_w)(uint8_t *_dst, const ptrdiff_t _dst_stride, | |
| 365 | const int16_t *src, const int16_t *diff_mv_x, const int16_t *diff_mv_y, | ||
| 366 | const int denom, const int wx, const int _ox) | ||
| 367 | { | ||
| 368 | 28664 | const int limit = (1 << FFMAX(13, BIT_DEPTH + 1)); ///< dILimit | |
| 369 | 28664 | pixel *dst = (pixel*)_dst; | |
| 370 | 28664 | const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel); | |
| 371 | 28664 | const int shift = denom + FFMAX(2, 14 - BIT_DEPTH); | |
| 372 | 28664 | const int offset = 1 << (shift - 1); | |
| 373 | 28664 | const int ox = _ox * (1 << (BIT_DEPTH - 8)); | |
| 374 | int16_t gradient_h[AFFINE_MIN_BLOCK_SIZE * AFFINE_MIN_BLOCK_SIZE]; | ||
| 375 | int16_t gradient_v[AFFINE_MIN_BLOCK_SIZE * AFFINE_MIN_BLOCK_SIZE]; | ||
| 376 | |||
| 377 | 28664 | FUNC(prof_grad_filter)(gradient_h, gradient_v, AFFINE_MIN_BLOCK_SIZE, src, MAX_PB_SIZE, AFFINE_MIN_BLOCK_SIZE, AFFINE_MIN_BLOCK_SIZE); | |
| 378 | |||
| 379 |
2/2✓ Branch 0 taken 57328 times.
✓ Branch 1 taken 14332 times.
|
143320 | for (int y = 0; y < AFFINE_MIN_BLOCK_SIZE; y++) { |
| 380 |
2/2✓ Branch 0 taken 229312 times.
✓ Branch 1 taken 57328 times.
|
573280 | for (int x = 0; x < AFFINE_MIN_BLOCK_SIZE; x++) { |
| 381 | 458624 | const int o = y * AFFINE_MIN_BLOCK_SIZE + x; | |
| 382 | 458624 | const int di = gradient_h[o] * diff_mv_x[o] + gradient_v[o] * diff_mv_y[o]; | |
| 383 | 458624 | const int val = src[x] + av_clip(di, -limit, limit - 1); | |
| 384 | 458624 | dst[x] = av_clip_pixel(((val * wx + offset) >> shift) + ox); | |
| 385 | } | ||
| 386 | 114656 | src += MAX_PB_SIZE; | |
| 387 | 114656 | dst += dst_stride; | |
| 388 | } | ||
| 389 | 28664 | } | |
| 390 | |||
| 391 | 10885408 | static void FUNC(derive_bdof_vx_vy)(const int16_t *_src0, const int16_t *_src1, | |
| 392 | const int pad_left, const int pad_top, const int pad_right, const int pad_bottom, | ||
| 393 | const int16_t **gradient_h, const int16_t **gradient_v, | ||
| 394 | int* vx, int* vy) | ||
| 395 | { | ||
| 396 | 10885408 | const int shift2 = 4; | |
| 397 | 10885408 | const int shift3 = 1; | |
| 398 | 10885408 | const int thres = 1 << 4; | |
| 399 | 10885408 | int sgx2 = 0, sgy2 = 0, sgxgy = 0, sgxdi = 0, sgydi = 0; | |
| 400 | |||
| 401 |
2/2✓ Branch 0 taken 32656224 times.
✓ Branch 1 taken 5442704 times.
|
76197856 | for (int y = -1; y < BDOF_MIN_BLOCK_SIZE + 1; y++) { |
| 402 |
8/8✓ Branch 0 taken 8345532 times.
✓ Branch 1 taken 24310692 times.
✓ Branch 2 taken 1390922 times.
✓ Branch 3 taken 6954610 times.
✓ Branch 4 taken 8345532 times.
✓ Branch 5 taken 24310692 times.
✓ Branch 6 taken 1390922 times.
✓ Branch 7 taken 6954610 times.
|
65312448 | const int dy = y + (pad_top && y < 0) - (pad_bottom && y == BDOF_MIN_BLOCK_SIZE); // we pad for the first and last row |
| 403 | 65312448 | const int16_t *src0 = _src0 + dy * MAX_PB_SIZE; | |
| 404 | 65312448 | const int16_t *src1 = _src1 + dy * MAX_PB_SIZE; | |
| 405 | |||
| 406 |
2/2✓ Branch 0 taken 195937344 times.
✓ Branch 1 taken 32656224 times.
|
457187136 | for (int x = -1; x < BDOF_MIN_BLOCK_SIZE + 1; x++) { |
| 407 |
8/8✓ Branch 0 taken 50402232 times.
✓ Branch 1 taken 145535112 times.
✓ Branch 2 taken 8400372 times.
✓ Branch 3 taken 42001860 times.
✓ Branch 4 taken 50402232 times.
✓ Branch 5 taken 145535112 times.
✓ Branch 6 taken 8400372 times.
✓ Branch 7 taken 42001860 times.
|
391874688 | const int dx = x + (pad_left && x < 0) - (pad_right && x == BDOF_MIN_BLOCK_SIZE); // we pad for the first and last col |
| 408 | 391874688 | const int diff = (src0[dx] >> shift2) - (src1[dx] >> shift2); | |
| 409 | 391874688 | const int idx = BDOF_BLOCK_SIZE * dy + dx; | |
| 410 | 391874688 | const int temph = (gradient_h[0][idx] + gradient_h[1][idx]) >> shift3; | |
| 411 | 391874688 | const int tempv = (gradient_v[0][idx] + gradient_v[1][idx]) >> shift3; | |
| 412 | |||
| 413 | 391874688 | sgx2 += FFABS(temph); | |
| 414 | 391874688 | sgy2 += FFABS(tempv); | |
| 415 |
2/2✓ Branch 0 taken 106288216 times.
✓ Branch 1 taken 89649128 times.
|
391874688 | sgxgy += VVC_SIGN(tempv) * temph; |
| 416 |
2/2✓ Branch 0 taken 112117480 times.
✓ Branch 1 taken 83819864 times.
|
391874688 | sgxdi += -VVC_SIGN(temph) * diff; |
| 417 |
2/2✓ Branch 0 taken 106288216 times.
✓ Branch 1 taken 89649128 times.
|
391874688 | sgydi += -VVC_SIGN(tempv) * diff; |
| 418 | } | ||
| 419 | } | ||
| 420 |
2/2✓ Branch 0 taken 5322931 times.
✓ Branch 1 taken 119773 times.
|
10885408 | *vx = sgx2 > 0 ? av_clip((sgxdi * (1 << 2)) >> av_log2(sgx2) , -thres + 1, thres - 1) : 0; |
| 421 |
2/2✓ Branch 0 taken 5328126 times.
✓ Branch 1 taken 114578 times.
|
10885408 | *vy = sgy2 > 0 ? av_clip(((sgydi * (1 << 2)) - ((*vx * sgxgy) >> 1)) >> av_log2(sgy2), -thres + 1, thres - 1) : 0; |
| 422 | 10885408 | } | |
| 423 | |||
| 424 | 10885408 | static void FUNC(apply_bdof_min_block)(pixel* dst, const ptrdiff_t dst_stride, const int16_t *src0, const int16_t *src1, | |
| 425 | const int16_t **gh, const int16_t **gv, const int vx, const int vy) | ||
| 426 | { | ||
| 427 | 10885408 | const int shift4 = 15 - BIT_DEPTH; | |
| 428 | 10885408 | const int offset4 = 1 << (shift4 - 1); | |
| 429 | |||
| 430 |
2/2✓ Branch 0 taken 21770816 times.
✓ Branch 1 taken 5442704 times.
|
54427040 | for (int y = 0; y < BDOF_MIN_BLOCK_SIZE; y++) { |
| 431 |
2/2✓ Branch 0 taken 87083264 times.
✓ Branch 1 taken 21770816 times.
|
217708160 | for (int x = 0; x < BDOF_MIN_BLOCK_SIZE; x++) { |
| 432 | 174166528 | const int idx = y * BDOF_BLOCK_SIZE + x; | |
| 433 | 174166528 | const int bdof_offset = vx * (gh[0][idx] - gh[1][idx]) + vy * (gv[0][idx] - gv[1][idx]); | |
| 434 | 174166528 | dst[x] = av_clip_pixel((src0[x] + offset4 + src1[x] + bdof_offset) >> shift4); | |
| 435 | } | ||
| 436 | 43541632 | dst += dst_stride; | |
| 437 | 43541632 | src0 += MAX_PB_SIZE; | |
| 438 | 43541632 | src1 += MAX_PB_SIZE; | |
| 439 | } | ||
| 440 | 10885408 | } | |
| 441 | |||
| 442 | 715154 | static void FUNC(apply_bdof)(uint8_t *_dst, const ptrdiff_t _dst_stride, const int16_t *_src0, const int16_t *_src1, | |
| 443 | const int block_w, const int block_h) | ||
| 444 | { | ||
| 445 | int16_t gradient_h[2][BDOF_BLOCK_SIZE * BDOF_BLOCK_SIZE]; | ||
| 446 | int16_t gradient_v[2][BDOF_BLOCK_SIZE * BDOF_BLOCK_SIZE]; | ||
| 447 | int vx, vy; | ||
| 448 | 715154 | const ptrdiff_t dst_stride = _dst_stride / sizeof(pixel); | |
| 449 | 715154 | pixel* dst = (pixel*)_dst; | |
| 450 | |||
| 451 | 715154 | FUNC(prof_grad_filter)(gradient_h[0], gradient_v[0], BDOF_BLOCK_SIZE, | |
| 452 | _src0, MAX_PB_SIZE, block_w, block_h); | ||
| 453 | 715154 | FUNC(prof_grad_filter)(gradient_h[1], gradient_v[1], BDOF_BLOCK_SIZE, | |
| 454 | _src1, MAX_PB_SIZE, block_w, block_h); | ||
| 455 | |||
| 456 |
2/2✓ Branch 0 taken 1400062 times.
✓ Branch 1 taken 357577 times.
|
3515278 | for (int y = 0; y < block_h; y += BDOF_MIN_BLOCK_SIZE) { |
| 457 |
2/2✓ Branch 0 taken 5442704 times.
✓ Branch 1 taken 1400062 times.
|
13685532 | for (int x = 0; x < block_w; x += BDOF_MIN_BLOCK_SIZE) { |
| 458 | 10885408 | const int16_t* src0 = _src0 + y * MAX_PB_SIZE + x; | |
| 459 | 10885408 | const int16_t* src1 = _src1 + y * MAX_PB_SIZE + x; | |
| 460 | 10885408 | pixel *d = dst + x; | |
| 461 | 10885408 | const int idx = BDOF_BLOCK_SIZE * y + x; | |
| 462 | 10885408 | const int16_t* gh[] = { gradient_h[0] + idx, gradient_h[1] + idx }; | |
| 463 | 10885408 | const int16_t* gv[] = { gradient_v[0] + idx, gradient_v[1] + idx }; | |
| 464 | 10885408 | FUNC(derive_bdof_vx_vy)(src0, src1, !x, !y, x + BDOF_MIN_BLOCK_SIZE == block_w, y + BDOF_MIN_BLOCK_SIZE == block_h, gh, gv, &vx, &vy); | |
| 465 | 10885408 | FUNC(apply_bdof_min_block)(d, dst_stride, src0, src1, gh, gv, vx, vy); | |
| 466 | } | ||
| 467 | 2800124 | dst += BDOF_MIN_BLOCK_SIZE * dst_stride; | |
| 468 | } | ||
| 469 | 715154 | } | |
| 470 | |||
| 471 | #define DMVR_FILTER(src, stride) \ | ||
| 472 | (filter[0] * src[x] + \ | ||
| 473 | filter[1] * src[x + stride]) | ||
| 474 | |||
| 475 | #define DMVR_FILTER2(filter, src0, src1) \ | ||
| 476 | (filter[0] * src0 + filter[1] * src1) | ||
| 477 | |||
| 478 | //8.5.3.2.2 Luma sample bilinear interpolation process | ||
| 479 | 2585898 | static void FUNC(dmvr)(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride, | |
| 480 | const int height, const intptr_t mx, const intptr_t my, const int width) | ||
| 481 | { | ||
| 482 | #if BIT_DEPTH != 10 | ||
| 483 | 36 | const pixel *src = (const pixel *)_src; | |
| 484 | 36 | const ptrdiff_t src_stride = _src_stride / sizeof(pixel); | |
| 485 | #if BIT_DEPTH > 10 | ||
| 486 | 18 | const int shift4 = BIT_DEPTH - 10; | |
| 487 | 18 | const int offset4 = 1 << (shift4 - 1); | |
| 488 | #define DMVR_SHIFT(s) (((s) + offset4) >> shift4) | ||
| 489 | #else | ||
| 490 | #define DMVR_SHIFT(s) ((s) << (10 - BIT_DEPTH)) | ||
| 491 | #endif // BIT_DEPTH > 10 | ||
| 492 | |||
| 493 |
2/2✓ Branch 0 taken 312 times.
✓ Branch 1 taken 18 times.
|
660 | for (int y = 0; y < height; y++) { |
| 494 |
2/2✓ Branch 0 taken 5280 times.
✓ Branch 1 taken 312 times.
|
11184 | for (int x = 0; x < width; x++) |
| 495 | 10560 | dst[x] = DMVR_SHIFT(src[x]); | |
| 496 | 624 | src += src_stride; | |
| 497 | 624 | dst += MAX_PB_SIZE; | |
| 498 | } | ||
| 499 | #undef DMVR_SHIFT | ||
| 500 | #else | ||
| 501 | 2585862 | av_image_copy_plane((uint8_t*)dst, sizeof(int16_t) * MAX_PB_SIZE, _src, _src_stride, | |
| 502 | width * sizeof(pixel), height); | ||
| 503 | #endif // BIT_DEPTH != 10 | ||
| 504 | 2585898 | } | |
| 505 | |||
| 506 | //8.5.3.2.2 Luma sample bilinear interpolation process | ||
| 507 | 308380 | static void FUNC(dmvr_h)(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride, | |
| 508 | const int height, const intptr_t mx, const intptr_t my, const int width) | ||
| 509 | { | ||
| 510 | 308380 | const pixel *src = (const pixel*)_src; | |
| 511 | 308380 | const ptrdiff_t src_stride = _src_stride / sizeof(pixel); | |
| 512 | 308380 | const int8_t *filter = ff_vvc_inter_luma_dmvr_filters[mx]; | |
| 513 | 308380 | const int shift1 = BIT_DEPTH - 6; | |
| 514 | 308380 | const int offset1 = 1 << (shift1 - 1); | |
| 515 | |||
| 516 |
2/2✓ Branch 0 taken 3050576 times.
✓ Branch 1 taken 154190 times.
|
6409532 | for (int y = 0; y < height; y++) { |
| 517 |
2/2✓ Branch 0 taken 60310720 times.
✓ Branch 1 taken 3050576 times.
|
126722592 | for (int x = 0; x < width; x++) |
| 518 | 120621440 | dst[x] = (DMVR_FILTER(src, 1) + offset1) >> shift1; | |
| 519 | 6101152 | src += src_stride; | |
| 520 | 6101152 | dst += MAX_PB_SIZE; | |
| 521 | } | ||
| 522 | 308380 | } | |
| 523 | |||
| 524 | //8.5.3.2.2 Luma sample bilinear interpolation process | ||
| 525 | 416760 | static void FUNC(dmvr_v)(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride, | |
| 526 | const int height, const intptr_t mx, const intptr_t my, const int width) | ||
| 527 | { | ||
| 528 | 416760 | const pixel *src = (pixel*)_src; | |
| 529 | 416760 | const ptrdiff_t src_stride = _src_stride / sizeof(pixel); | |
| 530 | 416760 | const int8_t *filter = ff_vvc_inter_luma_dmvr_filters[my]; | |
| 531 | 416760 | const int shift1 = BIT_DEPTH - 6; | |
| 532 | 416760 | const int offset1 = 1 << (shift1 - 1); | |
| 533 | |||
| 534 |
2/2✓ Branch 0 taken 4126784 times.
✓ Branch 1 taken 208380 times.
|
8670328 | for (int y = 0; y < height; y++) { |
| 535 |
2/2✓ Branch 0 taken 81890240 times.
✓ Branch 1 taken 4126784 times.
|
172034048 | for (int x = 0; x < width; x++) |
| 536 | 163780480 | dst[x] = (DMVR_FILTER(src, src_stride) + offset1) >> shift1; | |
| 537 | 8253568 | src += src_stride; | |
| 538 | 8253568 | dst += MAX_PB_SIZE; | |
| 539 | } | ||
| 540 | |||
| 541 | 416760 | } | |
| 542 | |||
| 543 | //8.5.3.2.2 Luma sample bilinear interpolation process | ||
| 544 | 1154354 | static void FUNC(dmvr_hv)(int16_t *dst, const uint8_t *_src, const ptrdiff_t _src_stride, | |
| 545 | const int height, const intptr_t mx, const intptr_t my, const int width) | ||
| 546 | { | ||
| 547 | int16_t tmp_array[MAX_PB_SIZE * 2]; | ||
| 548 | 1154354 | int16_t *tmp0 = tmp_array; | |
| 549 | 1154354 | int16_t *tmp1 = tmp_array + MAX_PB_SIZE; | |
| 550 | 1154354 | const pixel *src = (const pixel*)_src; | |
| 551 | 1154354 | const ptrdiff_t src_stride = _src_stride / sizeof(pixel); | |
| 552 | 1154354 | const int8_t *filter_x = ff_vvc_inter_luma_dmvr_filters[mx]; | |
| 553 | 1154354 | const int8_t *filter_y = ff_vvc_inter_luma_dmvr_filters[my]; | |
| 554 | 1154354 | const int shift1 = BIT_DEPTH - 6; | |
| 555 | 1154354 | const int offset1 = 1 << (shift1 - 1); | |
| 556 | 1154354 | const int shift2 = 4; | |
| 557 | 1154354 | const int offset2 = 1 << (shift2 - 1); | |
| 558 | |||
| 559 | 1154354 | src -= BILINEAR_EXTRA_BEFORE * src_stride; | |
| 560 |
2/2✓ Branch 0 taken 11383076 times.
✓ Branch 1 taken 577177 times.
|
23920506 | for (int x = 0; x < width; x++) |
| 561 | 22766152 | tmp0[x] = (DMVR_FILTER2(filter_x, src[x], src[x + 1]) + offset1) >> shift1; | |
| 562 | 1154354 | src += src_stride; | |
| 563 | |||
| 564 |
2/2✓ Branch 0 taken 11399980 times.
✓ Branch 1 taken 577177 times.
|
23954314 | for (int y = 1; y < height + BILINEAR_EXTRA; y++) { |
| 565 |
2/2✓ Branch 0 taken 224790320 times.
✓ Branch 1 taken 11399980 times.
|
472380600 | for (int x = 0; x < width; x++) { |
| 566 | 449580640 | tmp1[x] = (DMVR_FILTER2(filter_x, src[x], src[x + 1]) + offset1) >> shift1; | |
| 567 | 449580640 | dst[x] = (DMVR_FILTER2(filter_y, tmp0[x], tmp1[x]) + offset2) >> shift2; | |
| 568 | } | ||
| 569 | 22799960 | src += src_stride; | |
| 570 | 22799960 | dst += MAX_PB_SIZE; | |
| 571 | 22799960 | FFSWAP(int16_t *, tmp0, tmp1); | |
| 572 | } | ||
| 573 | 1154354 | } | |
| 574 | |||
| 575 | #define PEL_FUNC(dst, C, idx1, idx2, a) \ | ||
| 576 | do { \ | ||
| 577 | for (int w = 0; w < 7; w++) \ | ||
| 578 | inter->dst[C][w][idx1][idx2] = FUNC(a); \ | ||
| 579 | } while (0) \ | ||
| 580 | |||
| 581 | #define DIR_FUNCS(d, C, c) \ | ||
| 582 | PEL_FUNC(put_##d, C, 0, 0, put_##d##_pixels); \ | ||
| 583 | PEL_FUNC(put_##d, C, 0, 1, put_##d##_##c##_h); \ | ||
| 584 | PEL_FUNC(put_##d, C, 1, 0, put_##d##_##c##_v); \ | ||
| 585 | PEL_FUNC(put_##d, C, 1, 1, put_##d##_##c##_hv); \ | ||
| 586 | PEL_FUNC(put_##d##_w, C, 0, 0, put_##d##_w_pixels); \ | ||
| 587 | PEL_FUNC(put_##d##_w, C, 0, 1, put_##d##_##c##_w_h); \ | ||
| 588 | PEL_FUNC(put_##d##_w, C, 1, 0, put_##d##_##c##_w_v); \ | ||
| 589 | PEL_FUNC(put_##d##_w, C, 1, 1, put_##d##_##c##_w_hv); | ||
| 590 | |||
| 591 | #define FUNCS(C, c) \ | ||
| 592 | PEL_FUNC(put, C, 0, 0, put_pixels); \ | ||
| 593 | PEL_FUNC(put, C, 0, 1, put_##c##_h); \ | ||
| 594 | PEL_FUNC(put, C, 1, 0, put_##c##_v); \ | ||
| 595 | PEL_FUNC(put, C, 1, 1, put_##c##_hv); \ | ||
| 596 | DIR_FUNCS(uni, C, c); \ | ||
| 597 | |||
| 598 | 3084 | static void FUNC(ff_vvc_inter_dsp_init)(VVCInterDSPContext *const inter) | |
| 599 | { | ||
| 600 |
24/24✓ Branch 0 taken 10794 times.
✓ Branch 1 taken 1542 times.
✓ Branch 2 taken 10794 times.
✓ Branch 3 taken 1542 times.
✓ Branch 4 taken 10794 times.
✓ Branch 5 taken 1542 times.
✓ Branch 6 taken 10794 times.
✓ Branch 7 taken 1542 times.
✓ Branch 8 taken 10794 times.
✓ Branch 9 taken 1542 times.
✓ Branch 10 taken 10794 times.
✓ Branch 11 taken 1542 times.
✓ Branch 12 taken 10794 times.
✓ Branch 13 taken 1542 times.
✓ Branch 14 taken 10794 times.
✓ Branch 15 taken 1542 times.
✓ Branch 16 taken 10794 times.
✓ Branch 17 taken 1542 times.
✓ Branch 18 taken 10794 times.
✓ Branch 19 taken 1542 times.
✓ Branch 20 taken 10794 times.
✓ Branch 21 taken 1542 times.
✓ Branch 22 taken 10794 times.
✓ Branch 23 taken 1542 times.
|
262140 | FUNCS(LUMA, luma); |
| 601 |
24/24✓ Branch 0 taken 10794 times.
✓ Branch 1 taken 1542 times.
✓ Branch 2 taken 10794 times.
✓ Branch 3 taken 1542 times.
✓ Branch 4 taken 10794 times.
✓ Branch 5 taken 1542 times.
✓ Branch 6 taken 10794 times.
✓ Branch 7 taken 1542 times.
✓ Branch 8 taken 10794 times.
✓ Branch 9 taken 1542 times.
✓ Branch 10 taken 10794 times.
✓ Branch 11 taken 1542 times.
✓ Branch 12 taken 10794 times.
✓ Branch 13 taken 1542 times.
✓ Branch 14 taken 10794 times.
✓ Branch 15 taken 1542 times.
✓ Branch 16 taken 10794 times.
✓ Branch 17 taken 1542 times.
✓ Branch 18 taken 10794 times.
✓ Branch 19 taken 1542 times.
✓ Branch 20 taken 10794 times.
✓ Branch 21 taken 1542 times.
✓ Branch 22 taken 10794 times.
✓ Branch 23 taken 1542 times.
|
262140 | FUNCS(CHROMA, chroma); |
| 602 | |||
| 603 |
2/2✓ Branch 0 taken 10794 times.
✓ Branch 1 taken 1542 times.
|
24672 | for (int i = 0; i < FF_ARRAY_ELEMS(inter->put_scaled[LUMA]); i++) { |
| 604 | 21588 | inter->put_scaled[LUMA][i] = FUNC(put_luma_scaled); | |
| 605 | 21588 | inter->put_scaled[CHROMA][i] = FUNC(put_chroma_scaled); | |
| 606 | 21588 | inter->put_uni_scaled[LUMA][i] = FUNC(put_uni_luma_scaled); | |
| 607 | 21588 | inter->put_uni_scaled[CHROMA][i] = FUNC(put_uni_chroma_scaled); | |
| 608 | 21588 | inter->put_uni_w_scaled[LUMA][i] = FUNC(put_uni_luma_w_scaled); | |
| 609 | 21588 | inter->put_uni_w_scaled[CHROMA][i] = FUNC(put_uni_chroma_w_scaled); | |
| 610 | } | ||
| 611 | |||
| 612 | 3084 | inter->avg = FUNC(avg); | |
| 613 | 3084 | inter->w_avg = FUNC(w_avg); | |
| 614 | |||
| 615 | 3084 | inter->dmvr[0][0] = FUNC(dmvr); | |
| 616 | 3084 | inter->dmvr[0][1] = FUNC(dmvr_h); | |
| 617 | 3084 | inter->dmvr[1][0] = FUNC(dmvr_v); | |
| 618 | 3084 | inter->dmvr[1][1] = FUNC(dmvr_hv); | |
| 619 | |||
| 620 | 3084 | inter->put_ciip = FUNC(put_ciip); | |
| 621 | 3084 | inter->put_gpm = FUNC(put_gpm); | |
| 622 | |||
| 623 | 3084 | inter->fetch_samples = FUNC(fetch_samples); | |
| 624 | 3084 | inter->bdof_fetch_samples = FUNC(bdof_fetch_samples); | |
| 625 | 3084 | inter->apply_prof = FUNC(apply_prof); | |
| 626 | 3084 | inter->apply_prof_uni = FUNC(apply_prof_uni); | |
| 627 | 3084 | inter->apply_prof_uni_w = FUNC(apply_prof_uni_w); | |
| 628 | 3084 | inter->apply_bdof = FUNC(apply_bdof); | |
| 629 | 3084 | inter->sad = vvc_sad; | |
| 630 | 3084 | } | |
| 631 | |||
| 632 | #undef FUNCS | ||
| 633 | #undef PEL_FUNC | ||
| 634 | #undef DMVR_FUNCS | ||
| 635 |