1 // Copyright 2019 The libgav1 Authors
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 // http://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14
15 #include "src/dsp/cdef.h"
16
17 #include <algorithm>
18 #include <cassert>
19 #include <cstddef>
20 #include <cstdint>
21 #include <cstring>
22
23 #include "src/dsp/constants.h"
24 #include "src/dsp/dsp.h"
25 #include "src/utils/common.h"
26 #include "src/utils/constants.h"
27
28 namespace libgav1 {
29 namespace dsp {
30 namespace {
31
32 #include "src/dsp/cdef.inc"
33
34 // Silence unused function warnings when CdefDirection_C is obviated.
35 #if LIBGAV1_ENABLE_ALL_DSP_FUNCTIONS || \
36 !defined(LIBGAV1_Dsp8bpp_CdefDirection) || \
37 (LIBGAV1_MAX_BITDEPTH >= 10 && !defined(LIBGAV1_Dsp10bpp_CdefDirection))
38 constexpr int16_t kDivisionTable[] = {840, 420, 280, 210, 168, 140, 120, 105};
39
Square(int32_t x)40 int32_t Square(int32_t x) { return x * x; }
41
42 template <int bitdepth, typename Pixel>
CdefDirection_C(const void * LIBGAV1_RESTRICT const source,ptrdiff_t stride,uint8_t * LIBGAV1_RESTRICT const direction,int * LIBGAV1_RESTRICT const variance)43 void CdefDirection_C(const void* LIBGAV1_RESTRICT const source,
44 ptrdiff_t stride,
45 uint8_t* LIBGAV1_RESTRICT const direction,
46 int* LIBGAV1_RESTRICT const variance) {
47 assert(direction != nullptr);
48 assert(variance != nullptr);
49 const auto* src = static_cast<const Pixel*>(source);
50 stride /= sizeof(Pixel);
51 int32_t cost[8] = {};
52 // |partial| does not have to be int32_t for 8bpp. int16_t will suffice. We
53 // use int32_t to keep it simple since |cost| will have to be int32_t.
54 int32_t partial[8][15] = {};
55 for (int i = 0; i < 8; ++i) {
56 for (int j = 0; j < 8; ++j) {
57 const int x = (src[j] >> (bitdepth - 8)) - 128;
58 partial[0][i + j] += x;
59 partial[1][i + j / 2] += x;
60 partial[2][i] += x;
61 partial[3][3 + i - j / 2] += x;
62 partial[4][7 + i - j] += x;
63 partial[5][3 - i / 2 + j] += x;
64 partial[6][j] += x;
65 partial[7][i / 2 + j] += x;
66 }
67 src += stride;
68 }
69 for (int i = 0; i < 8; ++i) {
70 cost[2] += Square(partial[2][i]);
71 cost[6] += Square(partial[6][i]);
72 }
73 cost[2] *= kDivisionTable[7];
74 cost[6] *= kDivisionTable[7];
75 for (int i = 0; i < 7; ++i) {
76 cost[0] += (Square(partial[0][i]) + Square(partial[0][14 - i])) *
77 kDivisionTable[i];
78 cost[4] += (Square(partial[4][i]) + Square(partial[4][14 - i])) *
79 kDivisionTable[i];
80 }
81 cost[0] += Square(partial[0][7]) * kDivisionTable[7];
82 cost[4] += Square(partial[4][7]) * kDivisionTable[7];
83 for (int i = 1; i < 8; i += 2) {
84 for (int j = 0; j < 5; ++j) {
85 cost[i] += Square(partial[i][3 + j]);
86 }
87 cost[i] *= kDivisionTable[7];
88 for (int j = 0; j < 3; ++j) {
89 cost[i] += (Square(partial[i][j]) + Square(partial[i][10 - j])) *
90 kDivisionTable[2 * j + 1];
91 }
92 }
93 int32_t best_cost = 0;
94 *direction = 0;
95 for (int i = 0; i < 8; ++i) {
96 if (cost[i] > best_cost) {
97 best_cost = cost[i];
98 *direction = i;
99 }
100 }
101 *variance = (best_cost - cost[(*direction + 4) & 7]) >> 10;
102 }
103 #endif // LIBGAV1_ENABLE_ALL_DSP_FUNCTIONS ||
104 // !defined(LIBGAV1_Dsp8bpp_CdefDirection) ||
105 // (LIBGAV1_MAX_BITDEPTH >= 10 &&
106 // !defined(LIBGAV1_Dsp10bpp_CdefDirection))
107
108 // Silence unused function warnings when CdefFilter_C is obviated.
109 #if LIBGAV1_ENABLE_ALL_DSP_FUNCTIONS || \
110 !defined(LIBGAV1_Dsp8bpp_CdefFilters) || \
111 (LIBGAV1_MAX_BITDEPTH >= 10 && !defined(LIBGAV1_Dsp10bpp_CdefFilters))
112
Constrain(int diff,int threshold,int damping)113 int Constrain(int diff, int threshold, int damping) {
114 assert(threshold != 0);
115 damping = std::max(0, damping - FloorLog2(threshold));
116 const int sign = (diff < 0) ? -1 : 1;
117 return sign *
118 Clip3(threshold - (std::abs(diff) >> damping), 0, std::abs(diff));
119 }
120
121 // Filters the source block. It doesn't check whether the candidate pixel is
122 // inside the frame. However it requires the source input to be padded with a
123 // constant large value (kCdefLargeValue) if at the boundary.
124 template <int block_width, int bitdepth, typename Pixel,
125 bool enable_primary = true, bool enable_secondary = true>
CdefFilter_C(const uint16_t * LIBGAV1_RESTRICT src,const ptrdiff_t src_stride,const int block_height,const int primary_strength,const int secondary_strength,const int damping,const int direction,void * LIBGAV1_RESTRICT const dest,const ptrdiff_t dest_stride)126 void CdefFilter_C(const uint16_t* LIBGAV1_RESTRICT src,
127 const ptrdiff_t src_stride, const int block_height,
128 const int primary_strength, const int secondary_strength,
129 const int damping, const int direction,
130 void* LIBGAV1_RESTRICT const dest,
131 const ptrdiff_t dest_stride) {
132 static_assert(block_width == 4 || block_width == 8, "Invalid CDEF width.");
133 static_assert(enable_primary || enable_secondary, "");
134 assert(block_height == 4 || block_height == 8);
135 assert(direction >= 0 && direction <= 7);
136 constexpr int coeff_shift = bitdepth - 8;
137 // Section 5.9.19. CDEF params syntax.
138 assert(primary_strength >= 0 && primary_strength <= 15 << coeff_shift);
139 assert(secondary_strength >= 0 && secondary_strength <= 4 << coeff_shift &&
140 secondary_strength != 3 << coeff_shift);
141 assert(primary_strength != 0 || secondary_strength != 0);
142 // damping is decreased by 1 for chroma.
143 assert((damping >= 3 && damping <= 6 + coeff_shift) ||
144 (damping >= 2 && damping <= 5 + coeff_shift));
145 // When only primary_strength or secondary_strength are non-zero the number
146 // of pixels inspected (4 for primary_strength, 8 for secondary_strength) and
147 // the taps used don't exceed the amount the sum is
148 // descaled by (16) so we can skip tracking and clipping to the minimum and
149 // maximum value observed.
150 constexpr bool clipping_required = enable_primary && enable_secondary;
151 static constexpr int kCdefSecondaryTaps[2] = {kCdefSecondaryTap0,
152 kCdefSecondaryTap1};
153 auto* dst = static_cast<Pixel*>(dest);
154 const ptrdiff_t dst_stride = dest_stride / sizeof(Pixel);
155 int y = block_height;
156 do {
157 int x = 0;
158 do {
159 int16_t sum = 0;
160 const uint16_t pixel_value = src[x];
161 uint16_t max_value = pixel_value;
162 uint16_t min_value = pixel_value;
163 for (int k = 0; k < 2; ++k) {
164 static constexpr int signs[] = {-1, 1};
165 for (const int& sign : signs) {
166 if (enable_primary) {
167 const int dy = sign * kCdefDirections[direction][k][0];
168 const int dx = sign * kCdefDirections[direction][k][1];
169 const uint16_t value = src[dy * src_stride + dx + x];
170 // Note: the summation can ignore the condition check in SIMD
171 // implementation, because Constrain() will return 0 when
172 // value == kCdefLargeValue.
173 if (value != kCdefLargeValue) {
174 sum += Constrain(value - pixel_value, primary_strength, damping) *
175 kCdefPrimaryTaps[(primary_strength >> coeff_shift) & 1][k];
176 if (clipping_required) {
177 max_value = std::max(value, max_value);
178 min_value = std::min(value, min_value);
179 }
180 }
181 }
182
183 if (enable_secondary) {
184 static constexpr int offsets[] = {-2, 2};
185 for (const int& offset : offsets) {
186 const int dy = sign * kCdefDirections[direction + offset][k][0];
187 const int dx = sign * kCdefDirections[direction + offset][k][1];
188 const uint16_t value = src[dy * src_stride + dx + x];
189 // Note: the summation can ignore the condition check in SIMD
190 // implementation.
191 if (value != kCdefLargeValue) {
192 sum += Constrain(value - pixel_value, secondary_strength,
193 damping) *
194 kCdefSecondaryTaps[k];
195 if (clipping_required) {
196 max_value = std::max(value, max_value);
197 min_value = std::min(value, min_value);
198 }
199 }
200 }
201 }
202 }
203 }
204
205 const int offset = (8 + sum - (sum < 0)) >> 4;
206 if (clipping_required) {
207 dst[x] = static_cast<Pixel>(
208 Clip3(pixel_value + offset, min_value, max_value));
209 } else {
210 dst[x] = static_cast<Pixel>(pixel_value + offset);
211 }
212 } while (++x < block_width);
213
214 src += src_stride;
215 dst += dst_stride;
216 } while (--y != 0);
217 }
218 #endif // LIBGAV1_ENABLE_ALL_DSP_FUNCTIONS ||
219 // !defined(LIBGAV1_Dsp8bpp_CdefFilters) ||
220 // (LIBGAV1_MAX_BITDEPTH >= 10 &&
221 // !defined(LIBGAV1_Dsp10bpp_CdefFilters))
222
Init8bpp()223 void Init8bpp() {
224 Dsp* const dsp = dsp_internal::GetWritableDspTable(8);
225 assert(dsp != nullptr);
226 #if LIBGAV1_ENABLE_ALL_DSP_FUNCTIONS
227 dsp->cdef_direction = CdefDirection_C<8, uint8_t>;
228 dsp->cdef_filters[0][0] = CdefFilter_C<4, 8, uint8_t>;
229 dsp->cdef_filters[0][1] = CdefFilter_C<4, 8, uint8_t, /*enable_primary=*/true,
230 /*enable_secondary=*/false>;
231 dsp->cdef_filters[0][2] =
232 CdefFilter_C<4, 8, uint8_t, /*enable_primary=*/false>;
233 dsp->cdef_filters[1][0] = CdefFilter_C<8, 8, uint8_t>;
234 dsp->cdef_filters[1][1] = CdefFilter_C<8, 8, uint8_t, /*enable_primary=*/true,
235 /*enable_secondary=*/false>;
236 dsp->cdef_filters[1][2] =
237 CdefFilter_C<8, 8, uint8_t, /*enable_primary=*/false>;
238 #else // !LIBGAV1_ENABLE_ALL_DSP_FUNCTIONS
239 static_cast<void>(dsp);
240 #ifndef LIBGAV1_Dsp8bpp_CdefDirection
241 dsp->cdef_direction = CdefDirection_C<8, uint8_t>;
242 #endif
243 #ifndef LIBGAV1_Dsp8bpp_CdefFilters
244 dsp->cdef_filters[0][0] = CdefFilter_C<4, 8, uint8_t>;
245 dsp->cdef_filters[0][1] = CdefFilter_C<4, 8, uint8_t, /*enable_primary=*/true,
246 /*enable_secondary=*/false>;
247 dsp->cdef_filters[0][2] =
248 CdefFilter_C<4, 8, uint8_t, /*enable_primary=*/false>;
249 dsp->cdef_filters[1][0] = CdefFilter_C<8, 8, uint8_t>;
250 dsp->cdef_filters[1][1] = CdefFilter_C<8, 8, uint8_t, /*enable_primary=*/true,
251 /*enable_secondary=*/false>;
252 dsp->cdef_filters[1][2] =
253 CdefFilter_C<8, 8, uint8_t, /*enable_primary=*/false>;
254 #endif
255 #endif // LIBGAV1_ENABLE_ALL_DSP_FUNCTIONS
256 }
257
258 #if LIBGAV1_MAX_BITDEPTH >= 10
Init10bpp()259 void Init10bpp() {
260 Dsp* const dsp = dsp_internal::GetWritableDspTable(10);
261 assert(dsp != nullptr);
262 #if LIBGAV1_ENABLE_ALL_DSP_FUNCTIONS
263 dsp->cdef_direction = CdefDirection_C<10, uint16_t>;
264 dsp->cdef_filters[0][0] = CdefFilter_C<4, 10, uint16_t>;
265 dsp->cdef_filters[0][1] =
266 CdefFilter_C<4, 10, uint16_t, /*enable_primary=*/true,
267 /*enable_secondary=*/false>;
268 dsp->cdef_filters[0][2] =
269 CdefFilter_C<4, 10, uint16_t, /*enable_primary=*/false>;
270 dsp->cdef_filters[1][0] = CdefFilter_C<8, 10, uint16_t>;
271 dsp->cdef_filters[1][1] =
272 CdefFilter_C<8, 10, uint16_t, /*enable_primary=*/true,
273 /*enable_secondary=*/false>;
274 dsp->cdef_filters[1][2] =
275 CdefFilter_C<8, 10, uint16_t, /*enable_primary=*/false>;
276 #else // !LIBGAV1_ENABLE_ALL_DSP_FUNCTIONS
277 static_cast<void>(dsp);
278 #ifndef LIBGAV1_Dsp10bpp_CdefDirection
279 dsp->cdef_direction = CdefDirection_C<10, uint16_t>;
280 #endif
281 #ifndef LIBGAV1_Dsp10bpp_CdefFilters
282 dsp->cdef_filters[0][0] = CdefFilter_C<4, 10, uint16_t>;
283 dsp->cdef_filters[0][1] =
284 CdefFilter_C<4, 10, uint16_t, /*enable_primary=*/true,
285 /*enable_secondary=*/false>;
286 dsp->cdef_filters[0][2] =
287 CdefFilter_C<4, 10, uint16_t, /*enable_primary=*/false>;
288 dsp->cdef_filters[1][0] = CdefFilter_C<8, 10, uint16_t>;
289 dsp->cdef_filters[1][1] =
290 CdefFilter_C<8, 10, uint16_t, /*enable_primary=*/true,
291 /*enable_secondary=*/false>;
292 dsp->cdef_filters[1][2] =
293 CdefFilter_C<8, 10, uint16_t, /*enable_primary=*/false>;
294 #endif
295 #endif // LIBGAV1_ENABLE_ALL_DSP_FUNCTIONS
296 }
297 #endif
298
299 } // namespace
300
CdefInit_C()301 void CdefInit_C() {
302 Init8bpp();
303 #if LIBGAV1_MAX_BITDEPTH >= 10
304 Init10bpp();
305 #endif
306 }
307
308 } // namespace dsp
309 } // namespace libgav1
310