They are currently allocated separately per slice; allocating them jointly allows to avoid saving them in ff_update_duplicate_context(). The way it is done also avoids allocating ac_val for encoders that don't need it (e.g. H.263 or H.263+ with AIC). This entailed moving setting nb_slices to ff_mpv_init_context_frame() which is called from ff_mpv_common_frame_size_change(). The resultant nb_slices will always be one when called from a decoder using ff_mpv_common_frame_size_change().
Showing the changes to libavcodec (2 of 2 files). Show all
Unified · Split
| @@ -120,18 +120,6 @@ static av_cold int init_duplicate_context(MpegEncContext *s) · 119 unmodified lines | ||
| 120 | 120 | return AVERROR(ENOMEM); |
| 121 | 121 | s->block = s->blocks[0]; |
| 122 | 122 | |
| 123 | − if (s->out_format == FMT_H263) { | |
| 124 | − int mb_height = s->msmpeg4_version == MSMP4_VC1 ? | |
| 125 | − FFALIGN(s->mb_height, 2) : s->mb_height; | |
| 126 | − int y_size = s->b8_stride * (2 * mb_height + 1); | |
| 127 | − int c_size = s->mb_stride * (mb_height + 1); | |
| 128 | − int yc_size = y_size + 2 * c_size; | |
| 129 | − /* ac values */ | |
| 130 | − if (!FF_ALLOCZ_TYPED_ARRAY(s->ac_val_base, yc_size)) | |
| 131 | − return AVERROR(ENOMEM); | |
| 132 | − s->ac_val = s->ac_val_base + s->b8_stride + 1; | |
| 133 | − } | |
| 134 | − | |
| 135 | 123 | return 0; |
| 136 | 124 | } |
| 137 | 125 | |
| @@ -171,7 +159,6 @@ static av_cold void free_duplicate_context(MpegEncContext *s) · 33 unmodified lines | ||
| 171 | 159 | s->sc.linesize = 0; |
| 172 | 160 | |
| 173 | 161 | av_freep(&s->blocks); |
| 174 | − av_freep(&s->ac_val_base); | |
| 175 | 162 | s->block = NULL; |
| 176 | 163 | } |
| 177 | 164 | |
| @@ -192,7 +179,6 @@ static void backup_duplicate_context(MpegEncContext *bak, MpegEncContext *src) · 14 unmodified lines | ||
| 192 | 179 | COPY(block); |
| 193 | 180 | COPY(start_mb_y); |
| 194 | 181 | COPY(end_mb_y); |
| 195 | − COPY(ac_val_base); | |
| 196 | 182 | COPY(ac_val); |
| 197 | 183 | #undef COPY |
| 198 | 184 | } |
| @@ -245,15 +231,34 @@ static av_cold void free_buffer_pools(BufferPoolContext *pools) · 46 unmodified lines | ||
| 245 | 231 | |
| 246 | 232 | av_cold int ff_mpv_init_context_frame(MpegEncContext *s) |
| 247 | 233 | { |
| 234 | + int nb_slices = (HAVE_THREADS && | |
| 235 | + s->avctx->active_thread_type & FF_THREAD_SLICE) ? | |
| 236 | + s->avctx->thread_count : 1; | |
| 248 | 237 | BufferPoolContext *const pools = &s->buffer_pools; |
| 249 | 238 | int y_size, c_size, yc_size, i, mb_array_size, mv_table_size, x, y; |
| 250 | 239 | int mb_height; |
| 251 | 240 | |
| 241 | + if (s->encoding && s->avctx->slices) | |
| 242 | + nb_slices = s->avctx->slices; | |
| 243 | + | |
| 252 | 244 | if (s->codec_id == AV_CODEC_ID_MPEG2VIDEO && !s->progressive_sequence) |
| 253 | 245 | s->mb_height = (s->height + 31) / 32 * 2; |
| 254 | 246 | else |
| 255 | 247 | s->mb_height = (s->height + 15) / 16; |
| 256 | 248 | |
| 249 | + if (nb_slices > MAX_THREADS || (nb_slices > s->mb_height && s->mb_height)) { | |
| 250 | + int max_slices; | |
| 251 | + if (s->mb_height) | |
| 252 | + max_slices = FFMIN(MAX_THREADS, s->mb_height); | |
| 253 | + else | |
| 254 | + max_slices = MAX_THREADS; | |
| 255 | + av_log(s->avctx, AV_LOG_WARNING, "too many threads/slices (%d)," | |
| 256 | + " reducing to %d\n", nb_slices, max_slices); | |
| 257 | + nb_slices = max_slices; | |
| 258 | + } | |
| 259 | + | |
| 260 | + s->slice_context_count = nb_slices; | |
| 261 | + | |
| 257 | 262 | /* VC-1 can change from being progressive to interlaced on a per-frame |
| 258 | 263 | * basis. We therefore allocate certain buffers so big that they work |
| 259 | 264 | * in both instances. */ |
| @@ -332,6 +337,14 @@ av_cold int ff_mpv_init_context_frame(MpegEncContext *s) · 72 unmodified lines | ||
| 332 | 337 | } |
| 333 | 338 | |
| 334 | 339 | if (s->h263_pred || s->h263_aic || !s->encoding) { |
| 340 | + size_t allslice_yc_size = yc_size * (s->encoding ? nb_slices : 1); | |
| 341 | + if (s->out_format == FMT_H263) { | |
| 342 | + /* ac values */ | |
| 343 | + if (!FF_ALLOCZ_TYPED_ARRAY(s->ac_val_base, allslice_yc_size)) | |
| 344 | + return AVERROR(ENOMEM); | |
| 345 | + s->ac_val = s->ac_val_base + s->b8_stride + 1; | |
| 346 | + } | |
| 347 | + | |
| 335 | 348 | /* dc values */ |
| 336 | 349 | // MN: we need these for error resilience of intra-frames |
| 337 | 350 | // Allocating them unconditionally for decoders also means |
| @@ -381,14 +394,8 @@ av_cold int ff_mpv_init_context_frame(MpegEncContext *s) · 43 unmodified lines | ||
| 381 | 394 | */ |
| 382 | 395 | av_cold int ff_mpv_common_init(MpegEncContext *s) |
| 383 | 396 | { |
| 384 | − int nb_slices = (HAVE_THREADS && | |
| 385 | − s->avctx->active_thread_type & FF_THREAD_SLICE) ? | |
| 386 | − s->avctx->thread_count : 1; | |
| 387 | 397 | int ret; |
| 388 | 398 | |
| 389 | − if (s->encoding && s->avctx->slices) | |
| 390 | − nb_slices = s->avctx->slices; | |
| 391 | − | |
| 392 | 399 | if (s->avctx->pix_fmt == AV_PIX_FMT_NONE) { |
| 393 | 400 | av_log(s->avctx, AV_LOG_ERROR, |
| 394 | 401 | "decoding to AV_PIX_FMT_NONE is not supported.\n"); |
| @@ -411,20 +418,8 @@ av_cold int ff_mpv_common_init(MpegEncContext *s) · 16 unmodified lines | ||
| 411 | 418 | if ((ret = ff_mpv_init_context_frame(s))) |
| 412 | 419 | goto fail; |
| 413 | 420 | |
| 414 | − if (nb_slices > MAX_THREADS || (nb_slices > s->mb_height && s->mb_height)) { | |
| 415 | − int max_slices; | |
| 416 | − if (s->mb_height) | |
| 417 | − max_slices = FFMIN(MAX_THREADS, s->mb_height); | |
| 418 | − else | |
| 419 | − max_slices = MAX_THREADS; | |
| 420 | − av_log(s->avctx, AV_LOG_WARNING, "too many threads/slices (%d)," | |
| 421 | − " reducing to %d\n", nb_slices, max_slices); | |
| 422 | − nb_slices = max_slices; | |
| 423 | − } | |
| 424 | − | |
| 425 | 421 | s->context_initialized = 1; |
| 426 | 422 | s->thread_context[0] = s; |
| 427 | − s->slice_context_count = nb_slices; | |
| 428 | 423 | |
| 429 | 424 | // if (s->width && s->height) { |
| 430 | 425 | if (!s->encoding) { |
| @@ -450,6 +445,7 @@ av_cold void ff_mpv_free_context_frame(MpegEncContext *s) · 19 unmodified lines | ||
| 450 | 445 | for (int j = 0; j < 2; j++) |
| 451 | 446 | s->p_field_mv_table[i][j] = NULL; |
| 452 | 447 | |
| 448 | + av_freep(&s->ac_val_base); | |
| 453 | 449 | av_freep(&s->dc_val_base); |
| 454 | 450 | av_freep(&s->coded_block_base); |
| 455 | 451 | av_freep(&s->mbintra_table); |
| @@ -518,24 +518,36 @@ static av_cold int init_slice_buffers(MPVMainEncContext *const m) · 517 unmodified lines | ||
| 518 | 518 | static_assert(DCT_ERROR_SIZE * MAX_THREADS + ALIGN - 1 <= SIZE_MAX, |
| 519 | 519 | "Need checks for potential overflow."); |
| 520 | 520 | unsigned nb_slices = s->c.slice_context_count; |
| 521 | + char *dct_error = NULL; | |
| 521 | 522 | |
| 522 | − if (!m->noise_reduction) | |
| 523 | − return 0; | |
| 523 | + if (m->noise_reduction) { | |
| 524 | + if (!FF_ALLOCZ_TYPED_ARRAY(s->dct_offset, 2)) | |
| 525 | + return AVERROR(ENOMEM); | |
| 526 | + dct_error = av_mallocz(ALIGN - 1 + nb_slices * DCT_ERROR_SIZE); | |
| 527 | + if (!dct_error) | |
| 528 | + return AVERROR(ENOMEM); | |
| 529 | + m->dct_error_sum_base = dct_error; | |
| 530 | + dct_error += FFALIGN((uintptr_t)dct_error, ALIGN) - (uintptr_t)dct_error; | |
| 531 | + } | |
| 524 | 532 | |
| 525 | − if (!FF_ALLOCZ_TYPED_ARRAY(s->dct_offset, 2)) | |
| 526 | − return AVERROR(ENOMEM); | |
| 527 | − char *dct_error = av_mallocz(ALIGN - 1 + nb_slices * DCT_ERROR_SIZE); | |
| 528 | − if (!dct_error) | |
| 529 | − return AVERROR(ENOMEM); | |
| 530 | − m->dct_error_sum_base = dct_error; | |
| 531 | − dct_error += FFALIGN((uintptr_t)dct_error, ALIGN) - (uintptr_t)dct_error; | |
| 533 | + const int y_size = s->c.b8_stride * (2 * s->c.mb_height + 1); | |
| 534 | + const int c_size = s->c.mb_stride * (s->c.mb_height + 1); | |
| 535 | + const int yc_size = y_size + 2 * c_size; | |
| 536 | + ptrdiff_t offset = 0; | |
| 532 | 537 | |
| 533 | 538 | for (unsigned i = 0; i < nb_slices; ++i) { |
| 534 | 539 | MPVEncContext *const s2 = s->c.enc_contexts[i]; |
| 535 | 540 | |
| 536 | − s2->dct_offset = s->dct_offset; | |
| 537 | − s2->dct_error_sum = (void*)dct_error; | |
| 538 | − dct_error += DCT_ERROR_SIZE; | |
| 541 | + if (dct_error) { | |
| 542 | + s2->dct_offset = s->dct_offset; | |
| 543 | + s2->dct_error_sum = (void*)dct_error; | |
| 544 | + dct_error += DCT_ERROR_SIZE; | |
| 545 | + } | |
| 546 | + | |
| 547 | + if (s2->c.ac_val) { | |
| 548 | + s2->c.ac_val += offset; | |
| 549 | + offset += yc_size; | |
| 550 | + } | |
| 539 | 551 | } |
| 540 | 552 | return 0; |
| 541 | 553 | } |