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 <algorithm>
16 #include <cstdint>
17 #include <cstring>
18
19 #include "src/dsp/constants.h"
20 #include "src/obu_parser.h"
21 #include "src/symbol_decoder_context.h"
22 #include "src/tile.h"
23 #include "src/utils/array_2d.h"
24 #include "src/utils/block_parameters_holder.h"
25 #include "src/utils/common.h"
26 #include "src/utils/constants.h"
27 #include "src/utils/entropy_decoder.h"
28 #include "src/utils/segmentation.h"
29 #include "src/utils/stack.h"
30 #include "src/utils/types.h"
31
32 namespace libgav1 {
33 namespace {
34
35 constexpr uint8_t kMaxVariableTransformTreeDepth = 2;
36 // Max_Tx_Depth array from section 5.11.5 in the spec with the following
37 // modification: If the element is not zero, it is subtracted by one. That is
38 // the only way in which this array is being used.
39 constexpr int kTxDepthCdfIndex[kMaxBlockSizes] = {
40 0, 0, 1, 0, 0, 1, 2, 1, 1, 1, 2, 3, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3};
41
42 constexpr TransformSize kMaxTransformSizeRectangle[kMaxBlockSizes] = {
43 kTransformSize4x4, kTransformSize4x8, kTransformSize4x16,
44 kTransformSize8x4, kTransformSize8x8, kTransformSize8x16,
45 kTransformSize8x32, kTransformSize16x4, kTransformSize16x8,
46 kTransformSize16x16, kTransformSize16x32, kTransformSize16x64,
47 kTransformSize32x8, kTransformSize32x16, kTransformSize32x32,
48 kTransformSize32x64, kTransformSize64x16, kTransformSize64x32,
49 kTransformSize64x64, kTransformSize64x64, kTransformSize64x64,
50 kTransformSize64x64};
51
GetSquareTransformSize(uint8_t pixels)52 TransformSize GetSquareTransformSize(uint8_t pixels) {
53 switch (pixels) {
54 case 128:
55 case 64:
56 return kTransformSize64x64;
57 case 32:
58 return kTransformSize32x32;
59 case 16:
60 return kTransformSize16x16;
61 case 8:
62 return kTransformSize8x8;
63 default:
64 return kTransformSize4x4;
65 }
66 }
67
68 } // namespace
69
GetTopTransformWidth(const Block & block,int row4x4,int column4x4,bool ignore_skip)70 int Tile::GetTopTransformWidth(const Block& block, int row4x4, int column4x4,
71 bool ignore_skip) {
72 if (row4x4 == block.row4x4) {
73 if (!block.top_available[kPlaneY]) return 64;
74 const BlockParameters& bp_top =
75 *block_parameters_holder_.Find(row4x4 - 1, column4x4);
76 if ((ignore_skip || bp_top.skip) && bp_top.is_inter) {
77 return kBlockWidthPixels[bp_top.size];
78 }
79 }
80 return kTransformWidth[inter_transform_sizes_[row4x4 - 1][column4x4]];
81 }
82
GetLeftTransformHeight(const Block & block,int row4x4,int column4x4,bool ignore_skip)83 int Tile::GetLeftTransformHeight(const Block& block, int row4x4, int column4x4,
84 bool ignore_skip) {
85 if (column4x4 == block.column4x4) {
86 if (!block.left_available[kPlaneY]) return 64;
87 const BlockParameters& bp_left =
88 *block_parameters_holder_.Find(row4x4, column4x4 - 1);
89 if ((ignore_skip || bp_left.skip) && bp_left.is_inter) {
90 return kBlockHeightPixels[bp_left.size];
91 }
92 }
93 return kTransformHeight[inter_transform_sizes_[row4x4][column4x4 - 1]];
94 }
95
ReadFixedTransformSize(const Block & block)96 TransformSize Tile::ReadFixedTransformSize(const Block& block) {
97 BlockParameters& bp = *block.bp;
98 if (frame_header_.segmentation.lossless[bp.segment_id]) {
99 return kTransformSize4x4;
100 }
101 const TransformSize max_rect_tx_size = kMaxTransformSizeRectangle[block.size];
102 const bool allow_select = !bp.skip || !bp.is_inter;
103 if (block.size == kBlock4x4 || !allow_select ||
104 frame_header_.tx_mode != kTxModeSelect) {
105 return max_rect_tx_size;
106 }
107 const int max_tx_width = kTransformWidth[max_rect_tx_size];
108 const int max_tx_height = kTransformHeight[max_rect_tx_size];
109 const int top_width =
110 block.top_available[kPlaneY]
111 ? GetTopTransformWidth(block, block.row4x4, block.column4x4, true)
112 : 0;
113 const int left_height =
114 block.left_available[kPlaneY]
115 ? GetLeftTransformHeight(block, block.row4x4, block.column4x4, true)
116 : 0;
117 const auto context = static_cast<int>(top_width >= max_tx_width) +
118 static_cast<int>(left_height >= max_tx_height);
119 const int cdf_index = kTxDepthCdfIndex[block.size];
120 uint16_t* const cdf =
121 symbol_decoder_context_.tx_depth_cdf[cdf_index][context];
122 const int tx_depth = (cdf_index == 0)
123 ? static_cast<int>(reader_.ReadSymbol(cdf))
124 : reader_.ReadSymbol<3>(cdf);
125 assert(tx_depth < 3);
126 TransformSize tx_size = max_rect_tx_size;
127 if (tx_depth == 0) return tx_size;
128 tx_size = kSplitTransformSize[tx_size];
129 if (tx_depth == 1) return tx_size;
130 return kSplitTransformSize[tx_size];
131 }
132
ReadVariableTransformTree(const Block & block,int row4x4,int column4x4,TransformSize tx_size)133 void Tile::ReadVariableTransformTree(const Block& block, int row4x4,
134 int column4x4, TransformSize tx_size) {
135 const uint8_t pixels = std::max(block.width, block.height);
136 const TransformSize max_tx_size = GetSquareTransformSize(pixels);
137 const int context_delta = (kNumSquareTransformSizes - 1 -
138 TransformSizeToSquareTransformIndex(max_tx_size)) *
139 6;
140
141 // Branching factor is 4 and maximum depth is 2. So the maximum stack size
142 // necessary is (4 - 1) + 4 = 7.
143 Stack<TransformTreeNode, 7> stack;
144 stack.Push(TransformTreeNode(column4x4, row4x4, tx_size, 0));
145
146 do {
147 TransformTreeNode node = stack.Pop();
148 const int tx_width4x4 = kTransformWidth4x4[node.tx_size];
149 const int tx_height4x4 = kTransformHeight4x4[node.tx_size];
150 if (node.tx_size != kTransformSize4x4 &&
151 node.depth != kMaxVariableTransformTreeDepth) {
152 const auto top =
153 static_cast<int>(GetTopTransformWidth(block, node.y, node.x, false) <
154 kTransformWidth[node.tx_size]);
155 const auto left = static_cast<int>(
156 GetLeftTransformHeight(block, node.y, node.x, false) <
157 kTransformHeight[node.tx_size]);
158 const int context =
159 static_cast<int>(max_tx_size > kTransformSize8x8 &&
160 kTransformSizeSquareMax[node.tx_size] !=
161 max_tx_size) *
162 3 +
163 context_delta + top + left;
164 // tx_split.
165 if (reader_.ReadSymbol(symbol_decoder_context_.tx_split_cdf[context])) {
166 const TransformSize sub_tx_size = kSplitTransformSize[node.tx_size];
167 const int step_width4x4 = kTransformWidth4x4[sub_tx_size];
168 const int step_height4x4 = kTransformHeight4x4[sub_tx_size];
169 // The loops have to run in reverse order because we use a stack for
170 // DFS.
171 for (int i = tx_height4x4 - step_height4x4; i >= 0;
172 i -= step_height4x4) {
173 for (int j = tx_width4x4 - step_width4x4; j >= 0;
174 j -= step_width4x4) {
175 if (node.y + i >= frame_header_.rows4x4 ||
176 node.x + j >= frame_header_.columns4x4) {
177 continue;
178 }
179 stack.Push(TransformTreeNode(node.x + j, node.y + i, sub_tx_size,
180 node.depth + 1));
181 }
182 }
183 continue;
184 }
185 }
186 // tx_split is false.
187 for (int i = 0; i < tx_height4x4; ++i) {
188 static_assert(sizeof(TransformSize) == 1, "");
189 memset(&inter_transform_sizes_[node.y + i][node.x], node.tx_size,
190 tx_width4x4);
191 }
192 block_parameters_holder_.Find(node.y, node.x)->transform_size =
193 node.tx_size;
194 } while (!stack.Empty());
195 }
196
DecodeTransformSize(const Block & block)197 void Tile::DecodeTransformSize(const Block& block) {
198 BlockParameters& bp = *block.bp;
199 if (frame_header_.tx_mode == kTxModeSelect && block.size > kBlock4x4 &&
200 bp.is_inter && !bp.skip &&
201 !frame_header_.segmentation.lossless[bp.segment_id]) {
202 const TransformSize max_tx_size = kMaxTransformSizeRectangle[block.size];
203 const int tx_width4x4 = kTransformWidth4x4[max_tx_size];
204 const int tx_height4x4 = kTransformHeight4x4[max_tx_size];
205 for (int row = block.row4x4; row < block.row4x4 + block.height4x4;
206 row += tx_height4x4) {
207 for (int column = block.column4x4;
208 column < block.column4x4 + block.width4x4; column += tx_width4x4) {
209 ReadVariableTransformTree(block, row, column, max_tx_size);
210 }
211 }
212 } else {
213 bp.transform_size = ReadFixedTransformSize(block);
214 for (int row = block.row4x4; row < block.row4x4 + block.height4x4; ++row) {
215 static_assert(sizeof(TransformSize) == 1, "");
216 memset(&inter_transform_sizes_[row][block.column4x4], bp.transform_size,
217 block.width4x4);
218 }
219 }
220 }
221
222 } // namespace libgav1
223