pthread_frame.c 32.2 KB
Newer Older
1
/*
2
 * This file is part of FFmpeg.
3
 *
4
 * FFmpeg is free software; you can redistribute it and/or
5 6 7 8
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
9
 * FFmpeg is distributed in the hope that it will be useful,
10 11 12 13 14
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
15
 * License along with FFmpeg; if not, write to the Free Software
16 17 18 19 20 21 22 23 24 25 26
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

/**
 * @file
 * Frame multithreading support functions
 * @see doc/multithreading.txt
 */

#include "config.h"

27
#include <stdatomic.h>
28 29 30
#include <stdint.h>

#include "avcodec.h"
31
#include "hwaccel.h"
32 33 34
#include "internal.h"
#include "pthread_internal.h"
#include "thread.h"
35
#include "version.h"
36 37 38 39 40 41

#include "libavutil/avassert.h"
#include "libavutil/buffer.h"
#include "libavutil/common.h"
#include "libavutil/cpu.h"
#include "libavutil/frame.h"
42
#include "libavutil/internal.h"
43 44
#include "libavutil/log.h"
#include "libavutil/mem.h"
45
#include "libavutil/opt.h"
46
#include "libavutil/thread.h"
47

48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66
enum {
    ///< Set when the thread is awaiting a packet.
    STATE_INPUT_READY,
    ///< Set before the codec has called ff_thread_finish_setup().
    STATE_SETTING_UP,
    /**
     * Set when the codec calls get_buffer().
     * State is returned to STATE_SETTING_UP afterwards.
     */
    STATE_GET_BUFFER,
     /**
      * Set when the codec calls get_format().
      * State is returned to STATE_SETTING_UP afterwards.
      */
    STATE_GET_FORMAT,
    ///< Set after the codec has called ff_thread_finish_setup().
    STATE_SETUP_FINISHED,
};

67
/**
68
 * Context used by codec threads and stored in their AVCodecInternal thread_ctx.
69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85
 */
typedef struct PerThreadContext {
    struct FrameThreadContext *parent;

    pthread_t      thread;
    int            thread_init;
    pthread_cond_t input_cond;      ///< Used to wait for a new packet from the main thread.
    pthread_cond_t progress_cond;   ///< Used by child threads to wait for progress to change.
    pthread_cond_t output_cond;     ///< Used by the main thread to wait for frames to finish.

    pthread_mutex_t mutex;          ///< Mutex used to protect the contents of the PerThreadContext.
    pthread_mutex_t progress_mutex; ///< Mutex used to protect frame progress values and progress_cond.

    AVCodecContext *avctx;          ///< Context used to decode packets passed to this thread.

    AVPacket       avpkt;           ///< Input packet (for decoding) or output (for encoding).

86
    AVFrame *frame;                 ///< Output frame (for decoding) or input (for encoding).
87 88 89
    int     got_frame;              ///< The output of got_picture_ptr from the last avcodec_decode_video() call.
    int     result;                 ///< The result of the last codec decode/encode() call.

90
    atomic_int state;
91 92 93 94 95 96 97 98 99 100 101

    /**
     * Array of frames passed to ff_thread_release_buffer().
     * Frames are released after all threads referencing them are finished.
     */
    AVFrame *released_buffers;
    int  num_released_buffers;
    int      released_buffers_allocated;

    AVFrame *requested_frame;       ///< AVFrame the codec passed to get_buffer()
    int      requested_flags;       ///< flags passed to get_buffer() for requested_frame
102 103 104

    const enum AVPixelFormat *available_formats; ///< Format array for get_format()
    enum AVPixelFormat result_format;            ///< get_format() result
105 106

    int die;                        ///< Set when the thread should exit.
107 108

    int hwaccel_serializing;
109
    int async_serializing;
110 111 112
} PerThreadContext;

/**
113
 * Context stored in the client AVCodecInternal thread_ctx.
114 115 116 117 118 119
 */
typedef struct FrameThreadContext {
    PerThreadContext *threads;     ///< The contexts for each thread.
    PerThreadContext *prev_thread; ///< The last thread submit_packet() was called on.

    pthread_mutex_t buffer_mutex;  ///< Mutex used to protect get/release_buffer().
120 121 122 123 124
    /**
     * This lock is used for ensuring threads run in serial when hwaccel
     * is used.
     */
    pthread_mutex_t hwaccel_mutex;
125
    pthread_mutex_t async_mutex;
126 127
    pthread_cond_t async_cond;
    int async_lock;
128 129 130 131 132 133 134 135 136 137

    int next_decoding;             ///< The next context to submit a packet to.
    int next_finished;             ///< The next context to return output from.

    int delaying;                  /**<
                                    * Set for the first N packets, where N is the number of threads.
                                    * While it is set, ff_thread_en/decode_frame won't return any results.
                                    */
} FrameThreadContext;

138 139
#define THREAD_SAFE_CALLBACKS(avctx) \
((avctx)->thread_safe_callbacks || (avctx)->get_buffer2 == avcodec_default_get_buffer2)
140

141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158
static void async_lock(FrameThreadContext *fctx)
{
    pthread_mutex_lock(&fctx->async_mutex);
    while (fctx->async_lock)
        pthread_cond_wait(&fctx->async_cond, &fctx->async_mutex);
    fctx->async_lock = 1;
    pthread_mutex_unlock(&fctx->async_mutex);
}

static void async_unlock(FrameThreadContext *fctx)
{
    pthread_mutex_lock(&fctx->async_mutex);
    av_assert0(fctx->async_lock);
    fctx->async_lock = 0;
    pthread_cond_broadcast(&fctx->async_cond);
    pthread_mutex_unlock(&fctx->async_mutex);
}

159 160 161 162 163 164 165 166 167 168 169 170 171
/**
 * Codec worker thread.
 *
 * Automatically calls ff_thread_finish_setup() if the codec does
 * not provide an update_thread_context method, or if the codec returns
 * before calling it.
 */
static attribute_align_arg void *frame_worker_thread(void *arg)
{
    PerThreadContext *p = arg;
    AVCodecContext *avctx = p->avctx;
    const AVCodec *codec = avctx->codec;

172
    pthread_mutex_lock(&p->mutex);
173
    while (1) {
174 175
        while (atomic_load(&p->state) == STATE_INPUT_READY && !p->die)
            pthread_cond_wait(&p->input_cond, &p->mutex);
176

177
        if (p->die) break;
178

179
        if (!codec->update_thread_context && THREAD_SAFE_CALLBACKS(avctx))
180 181
            ff_thread_finish_setup(avctx);

182 183 184 185 186 187 188 189 190 191 192 193 194 195 196
        /* If a decoder supports hwaccel, then it must call ff_get_format().
         * Since that call must happen before ff_thread_finish_setup(), the
         * decoder is required to implement update_thread_context() and call
         * ff_thread_finish_setup() manually. Therefore the above
         * ff_thread_finish_setup() call did not happen and hwaccel_serializing
         * cannot be true here. */
        av_assert0(!p->hwaccel_serializing);

        /* if the previous thread uses hwaccel then we take the lock to ensure
         * the threads don't run concurrently */
        if (avctx->hwaccel) {
            pthread_mutex_lock(&p->parent->hwaccel_mutex);
            p->hwaccel_serializing = 1;
        }

197
        av_frame_unref(p->frame);
198
        p->got_frame = 0;
199
        p->result = codec->decode(avctx, p->frame, &p->got_frame, &p->avpkt);
200

201 202 203 204 205 206 207
        if ((p->result < 0 || !p->got_frame) && p->frame->buf[0]) {
            if (avctx->internal->allocate_progress)
                av_log(avctx, AV_LOG_ERROR, "A frame threaded decoder did not "
                       "free the frame on failure. This is a bug, please report it.\n");
            av_frame_unref(p->frame);
        }

208 209
        if (atomic_load(&p->state) == STATE_SETTING_UP)
            ff_thread_finish_setup(avctx);
210

211 212 213 214 215
        if (p->hwaccel_serializing) {
            p->hwaccel_serializing = 0;
            pthread_mutex_unlock(&p->parent->hwaccel_mutex);
        }

216 217
        if (p->async_serializing) {
            p->async_serializing = 0;
218 219

            async_unlock(p->parent);
220 221
        }

222
        pthread_mutex_lock(&p->progress_mutex);
223

224
        atomic_store(&p->state, STATE_INPUT_READY);
225

226
        pthread_cond_broadcast(&p->progress_cond);
227 228 229
        pthread_cond_signal(&p->output_cond);
        pthread_mutex_unlock(&p->progress_mutex);
    }
230
    pthread_mutex_unlock(&p->mutex);
231 232 233 234 235 236 237 238 239 240

    return NULL;
}

/**
 * Update the next thread's AVCodecContext with values from the reference thread's context.
 *
 * @param dst The destination context.
 * @param src The source context.
 * @param for_user 0 if the destination is a codec thread, 1 if the destination is the user's thread
241
 * @return 0 on success, negative error code on failure
242 243 244 245 246
 */
static int update_context_from_thread(AVCodecContext *dst, AVCodecContext *src, int for_user)
{
    int err = 0;

247
    if (dst != src && (for_user || !(av_codec_get_codec_descriptor(src)->props & AV_CODEC_PROP_INTRA_ONLY))) {
248
        dst->time_base = src->time_base;
249
        dst->framerate = src->framerate;
250 251 252
        dst->width     = src->width;
        dst->height    = src->height;
        dst->pix_fmt   = src->pix_fmt;
253
        dst->sw_pix_fmt = src->sw_pix_fmt;
254 255 256 257 258 259 260 261 262

        dst->coded_width  = src->coded_width;
        dst->coded_height = src->coded_height;

        dst->has_b_frames = src->has_b_frames;
        dst->idct_algo    = src->idct_algo;

        dst->bits_per_coded_sample = src->bits_per_coded_sample;
        dst->sample_aspect_ratio   = src->sample_aspect_ratio;
263 264
#if FF_API_AFD
FF_DISABLE_DEPRECATION_WARNINGS
265
        dst->dtg_active_format     = src->dtg_active_format;
266 267
FF_ENABLE_DEPRECATION_WARNINGS
#endif /* FF_API_AFD */
268 269 270 271 272 273 274 275 276 277 278 279 280 281 282

        dst->profile = src->profile;
        dst->level   = src->level;

        dst->bits_per_raw_sample = src->bits_per_raw_sample;
        dst->ticks_per_frame     = src->ticks_per_frame;
        dst->color_primaries     = src->color_primaries;

        dst->color_trc   = src->color_trc;
        dst->colorspace  = src->colorspace;
        dst->color_range = src->color_range;
        dst->chroma_sample_location = src->chroma_sample_location;

        dst->hwaccel = src->hwaccel;
        dst->hwaccel_context = src->hwaccel_context;
283 284 285 286 287

        dst->channels       = src->channels;
        dst->sample_rate    = src->sample_rate;
        dst->sample_fmt     = src->sample_fmt;
        dst->channel_layout = src->channel_layout;
288
        dst->internal->hwaccel_priv_data = src->internal->hwaccel_priv_data;
289 290 291 292 293 294 295 296 297 298 299

        if (!!dst->hw_frames_ctx != !!src->hw_frames_ctx ||
            (dst->hw_frames_ctx && dst->hw_frames_ctx->data != src->hw_frames_ctx->data)) {
            av_buffer_unref(&dst->hw_frames_ctx);

            if (src->hw_frames_ctx) {
                dst->hw_frames_ctx = av_buffer_ref(src->hw_frames_ctx);
                if (!dst->hw_frames_ctx)
                    return AVERROR(ENOMEM);
            }
        }
300 301

        dst->hwaccel_flags = src->hwaccel_flags;
302 303 304
    }

    if (for_user) {
305
        dst->delay       = src->thread_count - 1;
306 307
#if FF_API_CODED_FRAME
FF_DISABLE_DEPRECATION_WARNINGS
308
        dst->coded_frame = src->coded_frame;
309 310
FF_ENABLE_DEPRECATION_WARNINGS
#endif
311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335
    } else {
        if (dst->codec->update_thread_context)
            err = dst->codec->update_thread_context(dst, src);
    }

    return err;
}

/**
 * Update the next thread's AVCodecContext with values set by the user.
 *
 * @param dst The destination context.
 * @param src The source context.
 * @return 0 on success, negative error code on failure
 */
static int update_context_from_user(AVCodecContext *dst, AVCodecContext *src)
{
#define copy_fields(s, e) memcpy(&dst->s, &src->s, (char*)&dst->e - (char*)&dst->s);
    dst->flags          = src->flags;

    dst->draw_horiz_band= src->draw_horiz_band;
    dst->get_buffer2    = src->get_buffer2;

    dst->opaque   = src->opaque;
    dst->debug    = src->debug;
336
    dst->debug_mv = src->debug_mv;
337 338 339 340 341 342 343 344

    dst->slice_flags = src->slice_flags;
    dst->flags2      = src->flags2;

    copy_fields(skip_loop_filter, subtitle_header);

    dst->frame_number     = src->frame_number;
    dst->reordered_opaque = src->reordered_opaque;
345
    dst->thread_safe_callbacks = src->thread_safe_callbacks;
346 347 348

    if (src->slice_count && src->slice_offset) {
        if (dst->slice_count < src->slice_count) {
349 350 351 352
            int err = av_reallocp_array(&dst->slice_offset, src->slice_count,
                                        sizeof(*dst->slice_offset));
            if (err < 0)
                return err;
353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372
        }
        memcpy(dst->slice_offset, src->slice_offset,
               src->slice_count * sizeof(*dst->slice_offset));
    }
    dst->slice_count = src->slice_count;
    return 0;
#undef copy_fields
}

/// Releases the buffers that this decoding thread was the last user of.
static void release_delayed_buffers(PerThreadContext *p)
{
    FrameThreadContext *fctx = p->parent;

    while (p->num_released_buffers > 0) {
        AVFrame *f;

        pthread_mutex_lock(&fctx->buffer_mutex);

        // fix extended data in case the caller screwed it up
373 374
        av_assert0(p->avctx->codec_type == AVMEDIA_TYPE_VIDEO ||
                   p->avctx->codec_type == AVMEDIA_TYPE_AUDIO);
375 376 377 378 379 380 381 382
        f = &p->released_buffers[--p->num_released_buffers];
        f->extended_data = f->data;
        av_frame_unref(f);

        pthread_mutex_unlock(&fctx->buffer_mutex);
    }
}

383 384
static int submit_packet(PerThreadContext *p, AVCodecContext *user_avctx,
                         AVPacket *avpkt)
385 386 387 388
{
    FrameThreadContext *fctx = p->parent;
    PerThreadContext *prev_thread = fctx->prev_thread;
    const AVCodec *codec = p->avctx->codec;
389
    int ret;
390

391 392
    if (!avpkt->size && !(codec->capabilities & AV_CODEC_CAP_DELAY))
        return 0;
393 394 395

    pthread_mutex_lock(&p->mutex);

396 397 398 399 400 401
    ret = update_context_from_user(p->avctx, user_avctx);
    if (ret) {
        pthread_mutex_unlock(&p->mutex);
        return ret;
    }

402 403 404 405
    release_delayed_buffers(p);

    if (prev_thread) {
        int err;
406
        if (atomic_load(&prev_thread->state) == STATE_SETTING_UP) {
407
            pthread_mutex_lock(&prev_thread->progress_mutex);
408
            while (atomic_load(&prev_thread->state) == STATE_SETTING_UP)
409 410 411 412 413 414 415 416 417 418 419
                pthread_cond_wait(&prev_thread->progress_cond, &prev_thread->progress_mutex);
            pthread_mutex_unlock(&prev_thread->progress_mutex);
        }

        err = update_context_from_thread(p->avctx, prev_thread->avctx, 0);
        if (err) {
            pthread_mutex_unlock(&p->mutex);
            return err;
        }
    }

420
    av_packet_unref(&p->avpkt);
421 422 423 424 425 426
    ret = av_packet_ref(&p->avpkt, avpkt);
    if (ret < 0) {
        pthread_mutex_unlock(&p->mutex);
        av_log(p->avctx, AV_LOG_ERROR, "av_packet_ref() failed in submit_packet()\n");
        return ret;
    }
427

428
    atomic_store(&p->state, STATE_SETTING_UP);
429 430 431 432 433 434 435 436 437 438
    pthread_cond_signal(&p->input_cond);
    pthread_mutex_unlock(&p->mutex);

    /*
     * If the client doesn't have a thread-safe get_buffer(),
     * then decoding threads call back to the main thread,
     * and it calls back to the client here.
     */

    if (!p->avctx->thread_safe_callbacks && (
439
         p->avctx->get_format != avcodec_default_get_format ||
440
         p->avctx->get_buffer2 != avcodec_default_get_buffer2)) {
441
        while (atomic_load(&p->state) != STATE_SETUP_FINISHED && atomic_load(&p->state) != STATE_INPUT_READY) {
442
            int call_done = 1;
443
            pthread_mutex_lock(&p->progress_mutex);
444
            while (atomic_load(&p->state) == STATE_SETTING_UP)
445 446
                pthread_cond_wait(&p->progress_cond, &p->progress_mutex);

447
            switch (atomic_load_explicit(&p->state, memory_order_acquire)) {
448
            case STATE_GET_BUFFER:
449
                p->result = ff_get_buffer(p->avctx, p->requested_frame, p->requested_flags);
450 451
                break;
            case STATE_GET_FORMAT:
452
                p->result_format = ff_get_format(p->avctx, p->available_formats);
453 454 455 456 457 458
                break;
            default:
                call_done = 0;
                break;
            }
            if (call_done) {
459
                atomic_store(&p->state, STATE_SETTING_UP);
460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475
                pthread_cond_signal(&p->progress_cond);
            }
            pthread_mutex_unlock(&p->progress_mutex);
        }
    }

    fctx->prev_thread = p;
    fctx->next_decoding++;

    return 0;
}

int ff_thread_decode_frame(AVCodecContext *avctx,
                           AVFrame *picture, int *got_picture_ptr,
                           AVPacket *avpkt)
{
476
    FrameThreadContext *fctx = avctx->internal->thread_ctx;
477 478
    int finished = fctx->next_finished;
    PerThreadContext *p;
479
    int err;
480 481 482

    /* release the async lock, permitting blocked hwaccel threads to
     * go forward while we are in this function */
483
    async_unlock(fctx);
484 485 486 487 488 489

    /*
     * Submit a packet to the next decoding thread.
     */

    p = &fctx->threads[fctx->next_decoding];
490
    err = submit_packet(p, avctx, avpkt);
491 492
    if (err)
        goto finish;
493 494 495 496 497

    /*
     * If we're still receiving the initial packets, don't return a frame.
     */

498 499
    if (fctx->next_decoding > (avctx->thread_count-1-(avctx->codec_id == AV_CODEC_ID_FFV1)))
        fctx->delaying = 0;
500

501
    if (fctx->delaying) {
502
        *got_picture_ptr=0;
503
        if (avpkt->size) {
504
            err = avpkt->size;
505 506
            goto finish;
        }
507 508 509 510 511 512 513 514 515 516 517 518
    }

    /*
     * Return the next available frame from the oldest thread.
     * If we're at the end of the stream, then we have to skip threads that
     * didn't output a frame, because we don't want to accidentally signal
     * EOF (avpkt->size == 0 && *got_picture_ptr == 0).
     */

    do {
        p = &fctx->threads[finished++];

519
        if (atomic_load(&p->state) != STATE_INPUT_READY) {
520
            pthread_mutex_lock(&p->progress_mutex);
521
            while (atomic_load_explicit(&p->state, memory_order_relaxed) != STATE_INPUT_READY)
522 523 524 525
                pthread_cond_wait(&p->output_cond, &p->progress_mutex);
            pthread_mutex_unlock(&p->progress_mutex);
        }

526
        av_frame_move_ref(picture, p->frame);
527 528 529
        *got_picture_ptr = p->got_frame;
        picture->pkt_dts = p->avpkt.dts;

530 531 532
        if (p->result < 0)
            err = p->result;

533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550
        /*
         * A later call with avkpt->size == 0 may loop over all threads,
         * including this one, searching for a frame to return before being
         * stopped by the "finished != fctx->next_finished" condition.
         * Make sure we don't mistakenly return the same frame again.
         */
        p->got_frame = 0;

        if (finished >= avctx->thread_count) finished = 0;
    } while (!avpkt->size && !*got_picture_ptr && finished != fctx->next_finished);

    update_context_from_thread(avctx, p->avctx, 1);

    if (fctx->next_decoding >= avctx->thread_count) fctx->next_decoding = 0;

    fctx->next_finished = finished;

    /* return the size of the consumed packet if no error occurred */
551 552
    if (err >= 0)
        err = avpkt->size;
553
finish:
554
    async_lock(fctx);
555
    return err;
556 557 558 559 560
}

void ff_thread_report_progress(ThreadFrame *f, int n, int field)
{
    PerThreadContext *p;
561
    atomic_int *progress = f->progress ? (atomic_int*)f->progress->data : NULL;
562

563
    if (!progress ||
564
        atomic_load_explicit(&progress[field], memory_order_relaxed) >= n)
565
        return;
566

567
    p = f->owner[field]->internal->thread_ctx;
568

569
    pthread_mutex_lock(&p->progress_mutex);
570 571 572
    if (f->owner[field]->debug&FF_DEBUG_THREADS)
        av_log(f->owner[field], AV_LOG_DEBUG,
               "%p finished %d field %d\n", progress, n, field);
573

574
    atomic_store_explicit(&progress[field], n, memory_order_release);
575

576 577 578 579 580 581 582
    pthread_cond_broadcast(&p->progress_cond);
    pthread_mutex_unlock(&p->progress_mutex);
}

void ff_thread_await_progress(ThreadFrame *f, int n, int field)
{
    PerThreadContext *p;
583
    atomic_int *progress = f->progress ? (atomic_int*)f->progress->data : NULL;
584

585 586 587
    if (!progress ||
        atomic_load_explicit(&progress[field], memory_order_acquire) >= n)
        return;
588

589
    p = f->owner[field]->internal->thread_ctx;
590

591
    pthread_mutex_lock(&p->progress_mutex);
592 593 594
    if (f->owner[field]->debug&FF_DEBUG_THREADS)
        av_log(f->owner[field], AV_LOG_DEBUG,
               "thread awaiting %d field %d from %p\n", n, field, progress);
595
    while (atomic_load_explicit(&progress[field], memory_order_relaxed) < n)
596 597 598 599 600
        pthread_cond_wait(&p->progress_cond, &p->progress_mutex);
    pthread_mutex_unlock(&p->progress_mutex);
}

void ff_thread_finish_setup(AVCodecContext *avctx) {
601
    PerThreadContext *p = avctx->internal->thread_ctx;
602 603 604

    if (!(avctx->active_thread_type&FF_THREAD_FRAME)) return;

605 606 607 608 609
    if (avctx->hwaccel && !p->hwaccel_serializing) {
        pthread_mutex_lock(&p->parent->hwaccel_mutex);
        p->hwaccel_serializing = 1;
    }

610 611 612 613
    /* this assumes that no hwaccel calls happen before ff_thread_finish_setup() */
    if (avctx->hwaccel &&
        !(avctx->hwaccel->caps_internal & HWACCEL_CAP_ASYNC_SAFE)) {
        p->async_serializing = 1;
614 615

        async_lock(p->parent);
616 617
    }

618
    pthread_mutex_lock(&p->progress_mutex);
619
    if(atomic_load(&p->state) == STATE_SETUP_FINISHED){
620 621 622
        av_log(avctx, AV_LOG_WARNING, "Multiple ff_thread_finish_setup() calls\n");
    }

623 624
    atomic_store(&p->state, STATE_SETUP_FINISHED);

625 626 627 628 629 630 631 632 633
    pthread_cond_broadcast(&p->progress_cond);
    pthread_mutex_unlock(&p->progress_mutex);
}

/// Waits for all threads to finish.
static void park_frame_worker_threads(FrameThreadContext *fctx, int thread_count)
{
    int i;

634
    async_unlock(fctx);
635

636 637 638
    for (i = 0; i < thread_count; i++) {
        PerThreadContext *p = &fctx->threads[i];

639
        if (atomic_load(&p->state) != STATE_INPUT_READY) {
640
            pthread_mutex_lock(&p->progress_mutex);
641
            while (atomic_load(&p->state) != STATE_INPUT_READY)
642 643 644
                pthread_cond_wait(&p->output_cond, &p->progress_mutex);
            pthread_mutex_unlock(&p->progress_mutex);
        }
645
        p->got_frame = 0;
646
    }
647

648
    async_lock(fctx);
649 650 651 652
}

void ff_frame_thread_free(AVCodecContext *avctx, int thread_count)
{
653
    FrameThreadContext *fctx = avctx->internal->thread_ctx;
654 655 656 657 658 659
    const AVCodec *codec = avctx->codec;
    int i;

    park_frame_worker_threads(fctx, thread_count);

    if (fctx->prev_thread && fctx->prev_thread != fctx->threads)
660 661 662 663 664
        if (update_context_from_thread(fctx->threads->avctx, fctx->prev_thread->avctx, 0) < 0) {
            av_log(avctx, AV_LOG_ERROR, "Final thread update failed\n");
            fctx->prev_thread->avctx->internal->is_copy = fctx->threads->avctx->internal->is_copy;
            fctx->threads->avctx->internal->is_copy = 1;
        }
665 666 667 668 669

    for (i = 0; i < thread_count; i++) {
        PerThreadContext *p = &fctx->threads[i];

        pthread_mutex_lock(&p->mutex);
670
        p->die = 1;
671 672 673 674 675
        pthread_cond_signal(&p->input_cond);
        pthread_mutex_unlock(&p->mutex);

        if (p->thread_init)
            pthread_join(p->thread, NULL);
676
        p->thread_init=0;
677

678
        if (codec->close && p->avctx)
679 680 681
            codec->close(p->avctx);

        release_delayed_buffers(p);
682
        av_frame_free(&p->frame);
683 684 685 686 687 688 689 690 691 692
    }

    for (i = 0; i < thread_count; i++) {
        PerThreadContext *p = &fctx->threads[i];

        pthread_mutex_destroy(&p->mutex);
        pthread_mutex_destroy(&p->progress_mutex);
        pthread_cond_destroy(&p->input_cond);
        pthread_cond_destroy(&p->progress_cond);
        pthread_cond_destroy(&p->output_cond);
693
        av_packet_unref(&p->avpkt);
694 695
        av_freep(&p->released_buffers);

696
        if (i && p->avctx) {
697 698 699 700
            av_freep(&p->avctx->priv_data);
            av_freep(&p->avctx->slice_offset);
        }

701
        if (p->avctx) {
702
            av_freep(&p->avctx->internal);
703 704 705
            av_buffer_unref(&p->avctx->hw_frames_ctx);
        }

706 707 708 709 710
        av_freep(&p->avctx);
    }

    av_freep(&fctx->threads);
    pthread_mutex_destroy(&fctx->buffer_mutex);
711
    pthread_mutex_destroy(&fctx->hwaccel_mutex);
712
    pthread_mutex_destroy(&fctx->async_mutex);
713
    pthread_cond_destroy(&fctx->async_cond);
714

715
    av_freep(&avctx->internal->thread_ctx);
716 717 718 719

    if (avctx->priv_data && avctx->codec && avctx->codec->priv_class)
        av_opt_free(avctx->priv_data);
    avctx->codec = NULL;
720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735
}

int ff_frame_thread_init(AVCodecContext *avctx)
{
    int thread_count = avctx->thread_count;
    const AVCodec *codec = avctx->codec;
    AVCodecContext *src = avctx;
    FrameThreadContext *fctx;
    int i, err = 0;

#if HAVE_W32THREADS
    w32thread_init();
#endif

    if (!thread_count) {
        int nb_cpus = av_cpu_count();
736 737
        if ((avctx->debug & (FF_DEBUG_VIS_QP | FF_DEBUG_VIS_MB_TYPE)) || avctx->debug_mv)
            nb_cpus = 1;
738 739 740 741 742 743 744 745 746 747 748 749
        // use number of cores + 1 as thread count if there is more than one
        if (nb_cpus > 1)
            thread_count = avctx->thread_count = FFMIN(nb_cpus + 1, MAX_AUTO_THREADS);
        else
            thread_count = avctx->thread_count = 1;
    }

    if (thread_count <= 1) {
        avctx->active_thread_type = 0;
        return 0;
    }

750
    avctx->internal->thread_ctx = fctx = av_mallocz(sizeof(FrameThreadContext));
751 752
    if (!fctx)
        return AVERROR(ENOMEM);
753

754
    fctx->threads = av_mallocz_array(thread_count, sizeof(PerThreadContext));
755 756 757 758 759
    if (!fctx->threads) {
        av_freep(&avctx->internal->thread_ctx);
        return AVERROR(ENOMEM);
    }

760
    pthread_mutex_init(&fctx->buffer_mutex, NULL);
761
    pthread_mutex_init(&fctx->hwaccel_mutex, NULL);
762
    pthread_mutex_init(&fctx->async_mutex, NULL);
763
    pthread_cond_init(&fctx->async_cond, NULL);
764

765
    fctx->async_lock = 1;
766 767 768 769 770 771 772 773 774 775 776 777
    fctx->delaying = 1;

    for (i = 0; i < thread_count; i++) {
        AVCodecContext *copy = av_malloc(sizeof(AVCodecContext));
        PerThreadContext *p  = &fctx->threads[i];

        pthread_mutex_init(&p->mutex, NULL);
        pthread_mutex_init(&p->progress_mutex, NULL);
        pthread_cond_init(&p->input_cond, NULL);
        pthread_cond_init(&p->progress_cond, NULL);
        pthread_cond_init(&p->output_cond, NULL);

778 779
        p->frame = av_frame_alloc();
        if (!p->frame) {
780
            av_freep(&copy);
781 782 783 784
            err = AVERROR(ENOMEM);
            goto error;
        }

785 786 787 788 789 790 791 792 793 794
        p->parent = fctx;
        p->avctx  = copy;

        if (!copy) {
            err = AVERROR(ENOMEM);
            goto error;
        }

        *copy = *src;

795 796
        copy->internal = av_malloc(sizeof(AVCodecInternal));
        if (!copy->internal) {
797
            copy->priv_data = NULL;
798 799 800 801 802
            err = AVERROR(ENOMEM);
            goto error;
        }
        *copy->internal = *src->internal;
        copy->internal->thread_ctx = p;
803
        copy->internal->last_pkt_props = &p->avpkt;
804

805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826
        if (!i) {
            src = copy;

            if (codec->init)
                err = codec->init(copy);

            update_context_from_thread(avctx, copy, 1);
        } else {
            copy->priv_data = av_malloc(codec->priv_data_size);
            if (!copy->priv_data) {
                err = AVERROR(ENOMEM);
                goto error;
            }
            memcpy(copy->priv_data, src->priv_data, codec->priv_data_size);
            copy->internal->is_copy = 1;

            if (codec->init_thread_copy)
                err = codec->init_thread_copy(copy);
        }

        if (err) goto error;

827 828 829 830
        err = AVERROR(pthread_create(&p->thread, NULL, frame_worker_thread, p));
        p->thread_init= !err;
        if(!p->thread_init)
            goto error;
831 832 833 834 835 836 837 838 839 840 841 842 843
    }

    return 0;

error:
    ff_frame_thread_free(avctx, i+1);

    return err;
}

void ff_thread_flush(AVCodecContext *avctx)
{
    int i;
844
    FrameThreadContext *fctx = avctx->internal->thread_ctx;
845

846
    if (!fctx) return;
847 848 849 850 851 852 853 854 855 856 857 858 859 860

    park_frame_worker_threads(fctx, avctx->thread_count);
    if (fctx->prev_thread) {
        if (fctx->prev_thread != &fctx->threads[0])
            update_context_from_thread(fctx->threads[0].avctx, fctx->prev_thread->avctx, 0);
    }

    fctx->next_decoding = fctx->next_finished = 0;
    fctx->delaying = 1;
    fctx->prev_thread = NULL;
    for (i = 0; i < avctx->thread_count; i++) {
        PerThreadContext *p = &fctx->threads[i];
        // Make sure decode flush calls with size=0 won't return old frames
        p->got_frame = 0;
861
        av_frame_unref(p->frame);
862
        p->result = 0;
863 864

        release_delayed_buffers(p);
865 866 867

        if (avctx->codec->flush)
            avctx->codec->flush(p->avctx);
868 869 870
    }
}

871 872
int ff_thread_can_start_frame(AVCodecContext *avctx)
{
873
    PerThreadContext *p = avctx->internal->thread_ctx;
874
    if ((avctx->active_thread_type&FF_THREAD_FRAME) && atomic_load(&p->state) != STATE_SETTING_UP &&
875 876 877 878 879 880 881
        (avctx->codec->update_thread_context || !THREAD_SAFE_CALLBACKS(avctx))) {
        return 0;
    }
    return 1;
}

static int thread_get_buffer_internal(AVCodecContext *avctx, ThreadFrame *f, int flags)
882
{
883
    PerThreadContext *p = avctx->internal->thread_ctx;
884 885
    int err;

886
    f->owner[0] = f->owner[1] = avctx;
887

888 889
    ff_init_buffer_info(avctx, f->f);

890 891 892
    if (!(avctx->active_thread_type & FF_THREAD_FRAME))
        return ff_get_buffer(avctx, f->f, flags);

893
    if (atomic_load(&p->state) != STATE_SETTING_UP &&
894
        (avctx->codec->update_thread_context || !THREAD_SAFE_CALLBACKS(avctx))) {
895 896 897 898 899
        av_log(avctx, AV_LOG_ERROR, "get_buffer() cannot be called after ff_thread_finish_setup()\n");
        return -1;
    }

    if (avctx->internal->allocate_progress) {
900 901
        atomic_int *progress;
        f->progress = av_buffer_alloc(2 * sizeof(*progress));
902 903 904
        if (!f->progress) {
            return AVERROR(ENOMEM);
        }
905
        progress = (atomic_int*)f->progress->data;
906

907 908
        atomic_init(&progress[0], -1);
        atomic_init(&progress[1], -1);
909 910 911
    }

    pthread_mutex_lock(&p->parent->buffer_mutex);
912 913
    if (avctx->thread_safe_callbacks ||
        avctx->get_buffer2 == avcodec_default_get_buffer2) {
914 915
        err = ff_get_buffer(avctx, f->f, flags);
    } else {
916
        pthread_mutex_lock(&p->progress_mutex);
917 918
        p->requested_frame = f->f;
        p->requested_flags = flags;
919
        atomic_store_explicit(&p->state, STATE_GET_BUFFER, memory_order_release);
920
        pthread_cond_broadcast(&p->progress_cond);
921

922
        while (atomic_load(&p->state) != STATE_SETTING_UP)
923 924 925 926 927 928 929
            pthread_cond_wait(&p->progress_cond, &p->progress_mutex);

        err = p->result;

        pthread_mutex_unlock(&p->progress_mutex);

    }
930
    if (!THREAD_SAFE_CALLBACKS(avctx) && !avctx->codec->update_thread_context)
931 932 933 934 935 936 937 938 939
        ff_thread_finish_setup(avctx);
    if (err)
        av_buffer_unref(&f->progress);

    pthread_mutex_unlock(&p->parent->buffer_mutex);

    return err;
}

940 941 942
enum AVPixelFormat ff_thread_get_format(AVCodecContext *avctx, const enum AVPixelFormat *fmt)
{
    enum AVPixelFormat res;
943
    PerThreadContext *p = avctx->internal->thread_ctx;
944 945
    if (!(avctx->active_thread_type & FF_THREAD_FRAME) || avctx->thread_safe_callbacks ||
        avctx->get_format == avcodec_default_get_format)
946
        return ff_get_format(avctx, fmt);
947
    if (atomic_load(&p->state) != STATE_SETTING_UP) {
948 949 950 951 952
        av_log(avctx, AV_LOG_ERROR, "get_format() cannot be called after ff_thread_finish_setup()\n");
        return -1;
    }
    pthread_mutex_lock(&p->progress_mutex);
    p->available_formats = fmt;
953
    atomic_store(&p->state, STATE_GET_FORMAT);
954 955
    pthread_cond_broadcast(&p->progress_cond);

956
    while (atomic_load(&p->state) != STATE_SETTING_UP)
957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973
        pthread_cond_wait(&p->progress_cond, &p->progress_mutex);

    res = p->result_format;

    pthread_mutex_unlock(&p->progress_mutex);

    return res;
}

int ff_thread_get_buffer(AVCodecContext *avctx, ThreadFrame *f, int flags)
{
    int ret = thread_get_buffer_internal(avctx, f, flags);
    if (ret < 0)
        av_log(avctx, AV_LOG_ERROR, "thread_get_buffer() failed\n");
    return ret;
}

974 975
void ff_thread_release_buffer(AVCodecContext *avctx, ThreadFrame *f)
{
976
    PerThreadContext *p = avctx->internal->thread_ctx;
977 978 979 980
    FrameThreadContext *fctx;
    AVFrame *dst, *tmp;
    int can_direct_free = !(avctx->active_thread_type & FF_THREAD_FRAME) ||
                          avctx->thread_safe_callbacks                   ||
981
                          avctx->get_buffer2 == avcodec_default_get_buffer2;
982

983
    if (!f->f || !f->f->buf[0])
984 985 986 987 988 989
        return;

    if (avctx->debug & FF_DEBUG_BUFFERS)
        av_log(avctx, AV_LOG_DEBUG, "thread_release_buffer called on pic %p\n", f);

    av_buffer_unref(&f->progress);
990
    f->owner[0] = f->owner[1] = NULL;
991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016

    if (can_direct_free) {
        av_frame_unref(f->f);
        return;
    }

    fctx = p->parent;
    pthread_mutex_lock(&fctx->buffer_mutex);

    if (p->num_released_buffers + 1 >= INT_MAX / sizeof(*p->released_buffers))
        goto fail;
    tmp = av_fast_realloc(p->released_buffers, &p->released_buffers_allocated,
                          (p->num_released_buffers + 1) *
                          sizeof(*p->released_buffers));
    if (!tmp)
        goto fail;
    p->released_buffers = tmp;

    dst = &p->released_buffers[p->num_released_buffers];
    av_frame_move_ref(dst, f->f);

    p->num_released_buffers++;

fail:
    pthread_mutex_unlock(&fctx->buffer_mutex);
}