113 #define QUALITY_THRESHOLD 100
114 #define THRESHOLD_MULTIPLIER 0.6
118 int threshold,
int lambda,
int intra)
120 int count, y, x, i, j,
split, best_mean, best_score, best_count;
122 int block_sum[7] = { 0, 0, 0, 0, 0, 0 };
123 int w = 2 << (level + 2 >> 1);
124 int h = 2 << (level + 1 >> 1);
126 int16_t
block[7][256];
127 const int8_t *codebook_sum, *codebook;
128 const uint16_t(*mean_vlc)[2];
129 const uint8_t(*multistage_vlc)[2];
138 for (y = 0; y < h; y++) {
139 for (x = 0; x < w; x++) {
140 int v = src[x + y *
stride];
141 block[0][x + w * y] = v;
151 for (y = 0; y < h; y++) {
152 for (x = 0; x < w; x++) {
154 block[0][x + w * y] = v;
162 best_score -= (int)((
unsigned)block_sum[0] * block_sum[0] >> (level + 3));
163 best_mean = block_sum[0] + (size >> 1) >> (level + 3);
166 for (count = 1; count < 7; count++) {
167 int best_vector_score = INT_MAX;
168 int best_vector_sum = -999, best_vector_mean = -999;
169 const int stage = count - 1;
170 const int8_t *vector;
172 for (i = 0; i < 16; i++) {
173 int sum = codebook_sum[stage * 16 + i];
174 int sqr, diff, score;
176 vector = codebook + stage * size * 16 + i *
size;
178 diff = block_sum[stage] - sum;
179 score = sqr - (diff * (int64_t)diff >> (level + 3));
180 if (score < best_vector_score) {
181 int mean = diff + (size >> 1) >> (level + 3);
182 assert(mean > -300 && mean < 300);
183 mean = av_clip(mean, intra ? 0 : -256, 255);
184 best_vector_score = score;
185 best_vector[stage] = i;
186 best_vector_sum = sum;
187 best_vector_mean = mean;
190 assert(best_vector_mean != -999);
191 vector = codebook + stage * size * 16 + best_vector[stage] *
size;
192 for (j = 0; j <
size; j++)
193 block[stage + 1][j] = block[stage][j] - vector[j];
194 block_sum[stage + 1] = block_sum[stage] - best_vector_sum;
195 best_vector_score += lambda *
197 multistage_vlc[1 + count][1]
198 + mean_vlc[best_vector_mean][1]);
200 if (best_vector_score < best_score) {
201 best_score = best_vector_score;
203 best_mean = best_vector_mean;
209 if (best_score > threshold && level) {
211 int offset = level & 1 ? stride * h / 2 : w / 2;
214 for (i = level - 1; i >= 0; i--)
216 score +=
encode_block(s, src, ref, decoded, stride, level - 1,
217 threshold >> 1, lambda, intra);
218 score +=
encode_block(s, src + offset, ref + offset, decoded + offset,
219 stride, level - 1, threshold >> 1, lambda, intra);
222 if (score < best_score) {
226 for (i = level - 1; i >= 0; i--)
234 assert(best_mean >= 0 && best_mean < 256 || !intra);
235 assert(best_mean >= -256 && best_mean < 256);
236 assert(best_count >= 0 && best_count < 7);
237 assert(level < 4 || best_count == 0);
241 multistage_vlc[1 + best_count][1],
242 multistage_vlc[1 + best_count][0]);
244 mean_vlc[best_mean][0]);
246 for (i = 0; i < best_count; i++) {
247 assert(best_vector[i] >= 0 && best_vector[i] < 16);
251 for (y = 0; y < h; y++)
252 for (x = 0; x < w; x++)
253 decoded[x + y * stride] = src[x + y * stride] -
254 block[best_count][x + w * y] +
262 unsigned char *src_plane,
263 unsigned char *ref_plane,
264 unsigned char *decoded_plane,
270 int block_width, block_height;
279 for (level = 4; level >= 0; level--)
282 block_width = (width + 15) / 16;
283 block_height = (height + 15) / 16;
317 block_height * 2 + 2) *
318 2 *
sizeof(int16_t));
320 (block_height + 2) + 1) *
321 2 *
sizeof(int16_t));
340 for (y = 0; y < block_height; y++) {
344 for (i = 0; i < 16 && i + 16 * y <
height; i++) {
345 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
347 for (x = width; x < 16 * block_width; x++)
348 src[i * stride + x] = src[i * stride + x - 1];
350 for (; i < 16 && i + 16 * y < 16 * block_height; i++)
351 memcpy(&src[i * stride], &src[(i - 1) * stride],
354 for (x = 0; x < block_width; x++) {
370 for (y = 0; y < block_height; y++) {
371 for (i = 0; i < 16 && i + 16 * y <
height; i++) {
372 memcpy(&src[i * stride], &src_plane[(i + 16 * y) * src_stride],
374 for (x = width; x < 16 * block_width; x++)
375 src[i * stride + x] = src[i * stride + x - 1];
377 for (; i < 16 && i + 16 * y < 16 * block_height; i++)
378 memcpy(&src[i * stride], &src[(i - 1) * stride], 16 * block_width);
381 for (x = 0; x < block_width; x++) {
382 uint8_t reorder_buffer[3][6][7 * 32];
384 int offset = y * 16 * stride + x * 16;
385 uint8_t *decoded = decoded_plane + offset;
386 uint8_t *ref = ref_plane + offset;
387 int score[4] = { 0, 0, 0, 0 }, best;
403 for (i = 0; i < 6; i++)
409 score[0] = vlc[1] * lambda;
413 for (i = 0; i < 6; i++) {
424 int mx, my, pred_x, pred_y, dxy;
430 for (i = 0; i < 6; i++)
439 assert(mx >= -32 && mx <= 31);
440 assert(my >= -32 && my <= 31);
441 assert(pred_x >= -32 && pred_x <= 31);
442 assert(pred_y >= -32 && pred_y <= 31);
448 dxy = (mx & 1) + 2 * (my & 1);
456 decoded, stride, 5, 64, lambda, 0);
457 best = score[1] <= score[0];
460 score[2] = s->
dsp.
sse[0](
NULL, src + 16 * x, ref,
462 score[2] += vlc[1] * lambda;
463 if (score[2] < score[best] && mx == 0 && my == 0) {
466 for (i = 0; i < 6; i++)
473 for (i = 0; i < 6; i++) {
491 for (i = 5; i >= 0; i--)
518 for (i = 0; i < 3; i++) {
572 const AVFrame *pict,
int *got_packet)
606 for (i = 0; i < 3; i++)