// Allocate memory for lf row synchronization
void vp9_loop_filter_alloc(VP9LfSync *lf_sync, VP9_COMMON *cm, int rows,
int width) {
lf_sync->rows = rows;
#if CONFIG_MULTITHREAD
{
int i;
CHECK_MEM_ERROR(cm, lf_sync->mutex_,
vpx_malloc(sizeof(*lf_sync->mutex_) * rows));
for (i = 0; i < rows; ++i) {
pthread_mutex_init(&lf_sync->mutex_[i], NULL);
}
CHECK_MEM_ERROR(cm, lf_sync->cond_,
vpx_malloc(sizeof(*lf_sync->cond_) * rows));
for (i = 0; i < rows; ++i) {
pthread_cond_init(&lf_sync->cond_[i], NULL);
}
}
#endif // CONFIG_MULTITHREAD
CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
// Set up nsync.
lf_sync->sync_range = get_sync_range(width);
}
CYCLIC_REFRESH *vp9_cyclic_refresh_alloc(int mi_rows, int mi_cols) {
size_t last_coded_q_map_size;
size_t consec_zero_mv_size;
CYCLIC_REFRESH *const cr = vpx_calloc(1, sizeof(*cr));
if (cr == NULL)
return NULL;
cr->map = vpx_calloc(mi_rows * mi_cols, sizeof(*cr->map));
if (cr->map == NULL) {
vpx_free(cr);
return NULL;
}
last_coded_q_map_size = mi_rows * mi_cols * sizeof(*cr->last_coded_q_map);
cr->last_coded_q_map = vpx_malloc(last_coded_q_map_size);
if (cr->last_coded_q_map == NULL) {
vpx_free(cr);
return NULL;
}
assert(MAXQ <= 255);
memset(cr->last_coded_q_map, MAXQ, last_coded_q_map_size);
consec_zero_mv_size = mi_rows * mi_cols * sizeof(*cr->consec_zero_mv);
cr->consec_zero_mv = vpx_malloc(consec_zero_mv_size);
if (cr->consec_zero_mv == NULL) {
vpx_free(cr);
return NULL;
}
memset(cr->consec_zero_mv, 0, consec_zero_mv_size);
return cr;
}
// Allocate memory for row synchronization
void vp9_row_mt_sync_mem_alloc(VP9RowMTSync *row_mt_sync, VP9_COMMON *cm,
int rows) {
row_mt_sync->rows = rows;
#if CONFIG_MULTITHREAD
{
int i;
CHECK_MEM_ERROR(cm, row_mt_sync->mutex_,
vpx_malloc(sizeof(*row_mt_sync->mutex_) * rows));
if (row_mt_sync->mutex_) {
for (i = 0; i < rows; ++i) {
pthread_mutex_init(&row_mt_sync->mutex_[i], NULL);
}
}
CHECK_MEM_ERROR(cm, row_mt_sync->cond_,
vpx_malloc(sizeof(*row_mt_sync->cond_) * rows));
if (row_mt_sync->cond_) {
for (i = 0; i < rows; ++i) {
pthread_cond_init(&row_mt_sync->cond_[i], NULL);
}
}
}
#endif // CONFIG_MULTITHREAD
CHECK_MEM_ERROR(cm, row_mt_sync->cur_col,
vpx_malloc(sizeof(*row_mt_sync->cur_col) * rows));
// Set up nsync.
row_mt_sync->sync_range = 1;
}
void vp9_create_encoding_threads(VP9_COMP *cpi) {
VP9_COMMON * const cm = &cpi->common;
const VP9WorkerInterface * const winterface = vp9_get_worker_interface();
int i;
CHECK_MEM_ERROR(cm, cpi->enc_thread_hndl,
vpx_malloc(sizeof(*cpi->enc_thread_hndl) * cpi->max_threads));
for (i = 0; i < cpi->max_threads; ++i) {
VP9Worker * const worker = &cpi->enc_thread_hndl[i];
winterface->init(worker);
CHECK_MEM_ERROR(cm, worker->data1,
vpx_memalign(32, sizeof(thread_context)));
worker->data2 = NULL;
if (i < cpi->max_threads - 1 && !winterface->reset(worker)) {
vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
"Tile decoder thread creation failed");
}
}
// set row encoding hook
for (i = 0; i < cpi->max_threads; ++i) {
winterface->sync(&cpi->enc_thread_hndl[i]);
cpi->enc_thread_hndl[i].hook = (VP9WorkerHook) encoding_thread_process;
}
CHECK_MEM_ERROR(cm, cpi->cur_sb_col,
vpx_malloc(sizeof(*cpi->cur_sb_col) * cm->sb_rows));
// init cur sb col
vpx_memset(cpi->cur_sb_col, -1, (sizeof(*cpi->cur_sb_col) * cm->sb_rows));
// set up nsync (currently unused).
cpi->sync_range = get_sync_range(cpi->oxcf.width);
}
void vp8cx_create_encoder_threads(VP8_COMP *cpi)
{
cpi->b_multi_threaded = 0;
cpi->processor_core_count = 32; //vp8_get_proc_core_count();
CHECK_MEM_ERROR(cpi->tplist, vpx_malloc(sizeof(TOKENLIST) * cpi->common.mb_rows));
#if CONFIG_MULTITHREAD
if (cpi->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
{
int ithread;
if (cpi->oxcf.multi_threaded > cpi->processor_core_count)
cpi->encoding_thread_count = cpi->processor_core_count - 1;
else
cpi->encoding_thread_count = cpi->oxcf.multi_threaded - 1;
CHECK_MEM_ERROR(cpi->h_encoding_thread, vpx_malloc(sizeof(pthread_t) * cpi->encoding_thread_count));
CHECK_MEM_ERROR(cpi->h_event_mbrencoding, vpx_malloc(sizeof(sem_t) * cpi->encoding_thread_count));
CHECK_MEM_ERROR(cpi->mb_row_ei, vpx_memalign(32, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count));
vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * cpi->encoding_thread_count);
CHECK_MEM_ERROR(cpi->en_thread_data, vpx_malloc(sizeof(ENCODETHREAD_DATA) * cpi->encoding_thread_count));
//cpi->h_event_main = CreateEvent(NULL, FALSE, FALSE, NULL);
sem_init(&cpi->h_event_main, 0, 0);
cpi->b_multi_threaded = 1;
//printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n", (cpi->encoding_thread_count +1));
for (ithread = 0; ithread < cpi->encoding_thread_count; ithread++)
{
//cpi->h_event_mbrencoding[ithread] = CreateEvent(NULL, FALSE, FALSE, NULL);
sem_init(&cpi->h_event_mbrencoding[ithread], 0, 0);
cpi->en_thread_data[ithread].ithread = ithread;
cpi->en_thread_data[ithread].ptr1 = (void *)cpi;
cpi->en_thread_data[ithread].ptr2 = (void *)&cpi->mb_row_ei[ithread];
//printf(" call begin thread %d \n", ithread);
//cpi->h_encoding_thread[ithread] = (HANDLE)_beginthreadex(
// NULL, // security
// 0, // stksize
// thread_encoding_proc,
// (&cpi->en_thread_data[ithread]), // Thread data
// 0,
// NULL);
pthread_create(&cpi->h_encoding_thread[ithread], 0, thread_encoding_proc, (&cpi->en_thread_data[ithread]));
}
}
#endif
}
void vp9_dec_alloc_row_mt_mem(RowMTWorkerData *row_mt_worker_data,
VP9_COMMON *cm, int num_sbs, int max_threads,
int num_jobs) {
int plane;
const size_t dqcoeff_size = (num_sbs << DQCOEFFS_PER_SB_LOG2) *
sizeof(*row_mt_worker_data->dqcoeff[0]);
row_mt_worker_data->num_jobs = num_jobs;
#if CONFIG_MULTITHREAD
{
int i;
CHECK_MEM_ERROR(
cm, row_mt_worker_data->recon_sync_mutex,
vpx_malloc(sizeof(*row_mt_worker_data->recon_sync_mutex) * num_jobs));
if (row_mt_worker_data->recon_sync_mutex) {
for (i = 0; i < num_jobs; ++i) {
pthread_mutex_init(&row_mt_worker_data->recon_sync_mutex[i], NULL);
}
}
CHECK_MEM_ERROR(
cm, row_mt_worker_data->recon_sync_cond,
vpx_malloc(sizeof(*row_mt_worker_data->recon_sync_cond) * num_jobs));
if (row_mt_worker_data->recon_sync_cond) {
for (i = 0; i < num_jobs; ++i) {
pthread_cond_init(&row_mt_worker_data->recon_sync_cond[i], NULL);
}
}
}
#endif
row_mt_worker_data->num_sbs = num_sbs;
for (plane = 0; plane < 3; ++plane) {
CHECK_MEM_ERROR(cm, row_mt_worker_data->dqcoeff[plane],
vpx_memalign(16, dqcoeff_size));
memset(row_mt_worker_data->dqcoeff[plane], 0, dqcoeff_size);
CHECK_MEM_ERROR(cm, row_mt_worker_data->eob[plane],
vpx_calloc(num_sbs << EOBS_PER_SB_LOG2,
sizeof(*row_mt_worker_data->eob[plane])));
}
CHECK_MEM_ERROR(cm, row_mt_worker_data->partition,
vpx_calloc(num_sbs * PARTITIONS_PER_SB,
sizeof(*row_mt_worker_data->partition)));
CHECK_MEM_ERROR(cm, row_mt_worker_data->recon_map,
vpx_calloc(num_sbs, sizeof(*row_mt_worker_data->recon_map)));
// allocate memory for thread_data
if (row_mt_worker_data->thread_data == NULL) {
const size_t thread_size =
max_threads * sizeof(*row_mt_worker_data->thread_data);
CHECK_MEM_ERROR(cm, row_mt_worker_data->thread_data,
vpx_memalign(32, thread_size));
}
}
static void create_enc_workers(VP9_COMP *cpi, int num_workers) {
VP9_COMMON *const cm = &cpi->common;
const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
int i;
// Only run once to create threads and allocate thread data.
if (cpi->num_workers == 0) {
int allocated_workers = num_workers;
// While using SVC, we need to allocate threads according to the highest
// resolution. When row based multithreading is enabled, it is OK to
// allocate more threads than the number of max tile columns.
if (cpi->use_svc && !cpi->row_mt) {
int max_tile_cols = get_max_tile_cols(cpi);
allocated_workers = VPXMIN(cpi->oxcf.max_threads, max_tile_cols);
}
CHECK_MEM_ERROR(cm, cpi->workers,
vpx_malloc(allocated_workers * sizeof(*cpi->workers)));
CHECK_MEM_ERROR(cm, cpi->tile_thr_data,
vpx_calloc(allocated_workers, sizeof(*cpi->tile_thr_data)));
for (i = 0; i < allocated_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *thread_data = &cpi->tile_thr_data[i];
++cpi->num_workers;
winterface->init(worker);
if (i < allocated_workers - 1) {
thread_data->cpi = cpi;
// Allocate thread data.
CHECK_MEM_ERROR(cm, thread_data->td,
vpx_memalign(32, sizeof(*thread_data->td)));
vp9_zero(*thread_data->td);
// Set up pc_tree.
thread_data->td->leaf_tree = NULL;
thread_data->td->pc_tree = NULL;
vp9_setup_pc_tree(cm, thread_data->td);
// Allocate frame counters in thread data.
CHECK_MEM_ERROR(cm, thread_data->td->counts,
vpx_calloc(1, sizeof(*thread_data->td->counts)));
// Create threads
if (!winterface->reset(worker))
vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
"Tile encoder thread creation failed");
} else {
// Main thread acts as a worker and uses the thread data in cpi.
thread_data->cpi = cpi;
thread_data->td = &cpi->td;
}
winterface->sync(worker);
}
}
}
struct frame_dec_param *frame_dec_param_get(struct task *tsk) {
struct frame_dec_param *param;
param = vpx_malloc(sizeof(*param));
if (!param) {
return NULL;
}
tsk->priv = param;
tsk->dtor = task_dtor;
return param;
}
static vpx_codec_err_t init_decoder(vpx_codec_alg_priv_t *ctx) {
int i;
const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
ctx->last_show_frame = -1;
ctx->next_submit_worker_id = 0;
ctx->last_submit_worker_id = 0;
ctx->next_output_worker_id = 0;
ctx->frame_cache_read = 0;
ctx->frame_cache_write = 0;
ctx->num_cache_frames = 0;
ctx->need_resync = 1;
ctx->num_frame_workers =
(ctx->frame_parallel_decode == 1) ? ctx->cfg.threads: 1;
if (ctx->num_frame_workers > MAX_DECODE_THREADS)
ctx->num_frame_workers = MAX_DECODE_THREADS;
ctx->available_threads = ctx->num_frame_workers;
ctx->flushed = 0;
ctx->buffer_pool = (BufferPool *)vpx_calloc(1, sizeof(BufferPool));
if (ctx->buffer_pool == NULL)
return VPX_CODEC_MEM_ERROR;
#if CONFIG_MULTITHREAD
if (pthread_mutex_init(&ctx->buffer_pool->pool_mutex, NULL)) {
set_error_detail(ctx, "Failed to allocate buffer pool mutex");
return VPX_CODEC_MEM_ERROR;
}
#endif
ctx->frame_workers = (VPxWorker *)
vpx_malloc(ctx->num_frame_workers * sizeof(*ctx->frame_workers));
if (ctx->frame_workers == NULL) {
set_error_detail(ctx, "Failed to allocate frame_workers");
return VPX_CODEC_MEM_ERROR;
}
for (i = 0; i < ctx->num_frame_workers; ++i) {
VPxWorker *const worker = &ctx->frame_workers[i];
FrameWorkerData *frame_worker_data = NULL;
winterface->init(worker);
worker->data1 = vpx_memalign(32, sizeof(FrameWorkerData));
if (worker->data1 == NULL) {
set_error_detail(ctx, "Failed to allocate frame_worker_data");
return VPX_CODEC_MEM_ERROR;
}
frame_worker_data = (FrameWorkerData *)worker->data1;
frame_worker_data->pbi = vp9_decoder_create(ctx->buffer_pool);
if (frame_worker_data->pbi == NULL) {
set_error_detail(ctx, "Failed to allocate frame_worker_data");
return VPX_CODEC_MEM_ERROR;
}
frame_worker_data->pbi->frame_worker_owner = worker;
frame_worker_data->worker_id = i;
frame_worker_data->scratch_buffer = NULL;
frame_worker_data->scratch_buffer_size = 0;
frame_worker_data->frame_context_ready = 0;
frame_worker_data->received_frame = 0;
#if CONFIG_MULTITHREAD
if (pthread_mutex_init(&frame_worker_data->stats_mutex, NULL)) {
set_error_detail(ctx, "Failed to allocate frame_worker_data mutex");
return VPX_CODEC_MEM_ERROR;
}
if (pthread_cond_init(&frame_worker_data->stats_cond, NULL)) {
set_error_detail(ctx, "Failed to allocate frame_worker_data cond");
return VPX_CODEC_MEM_ERROR;
}
#endif
// If decoding in serial mode, FrameWorker thread could create tile worker
// thread or loopfilter thread.
frame_worker_data->pbi->max_threads =
(ctx->frame_parallel_decode == 0) ? ctx->cfg.threads : 0;
frame_worker_data->pbi->inv_tile_order = ctx->invert_tile_order;
frame_worker_data->pbi->frame_parallel_decode = ctx->frame_parallel_decode;
frame_worker_data->pbi->common.frame_parallel_decode =
ctx->frame_parallel_decode;
worker->hook = (VPxWorkerHook)frame_worker_hook;
if (!winterface->reset(worker)) {
set_error_detail(ctx, "Frame Worker thread creation failed");
return VPX_CODEC_MEM_ERROR;
}
}
// If postprocessing was enabled by the application and a
// configuration has not been provided, default it.
if (!ctx->postproc_cfg_set &&
(ctx->base.init_flags & VPX_CODEC_USE_POSTPROC))
set_default_ppflags(&ctx->postproc_cfg);
init_buffer_callbacks(ctx);
return VPX_CODEC_OK;
}
int vp8cx_create_encoder_threads(VP8_COMP *cpi)
{
const VP8_COMMON * cm = &cpi->common;
cpi->b_multi_threaded = 0;
cpi->encoding_thread_count = 0;
cpi->b_lpf_running = 0;
if (cm->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
{
int ithread;
int th_count = cpi->oxcf.multi_threaded - 1;
int rc = 0;
/* don't allocate more threads than cores available */
if (cpi->oxcf.multi_threaded > cm->processor_core_count)
th_count = cm->processor_core_count - 1;
/* we have th_count + 1 (main) threads processing one row each */
/* no point to have more threads than the sync range allows */
if(th_count > ((cm->mb_cols / cpi->mt_sync_range) - 1))
{
th_count = (cm->mb_cols / cpi->mt_sync_range) - 1;
}
if(th_count == 0)
return 0;
CHECK_MEM_ERROR(cpi->h_encoding_thread,
vpx_malloc(sizeof(pthread_t) * th_count));
CHECK_MEM_ERROR(cpi->h_event_start_encoding,
vpx_malloc(sizeof(sem_t) * th_count));
CHECK_MEM_ERROR(cpi->mb_row_ei,
vpx_memalign(32, sizeof(MB_ROW_COMP) * th_count));
vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * th_count);
CHECK_MEM_ERROR(cpi->en_thread_data,
vpx_malloc(sizeof(ENCODETHREAD_DATA) * th_count));
sem_init(&cpi->h_event_end_encoding, 0, 0);
cpi->b_multi_threaded = 1;
cpi->encoding_thread_count = th_count;
/*
printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n",
(cpi->encoding_thread_count +1));
*/
for (ithread = 0; ithread < th_count; ithread++)
{
ENCODETHREAD_DATA *ethd = &cpi->en_thread_data[ithread];
/* Setup block ptrs and offsets */
vp8_setup_block_ptrs(&cpi->mb_row_ei[ithread].mb);
vp8_setup_block_dptrs(&cpi->mb_row_ei[ithread].mb.e_mbd);
sem_init(&cpi->h_event_start_encoding[ithread], 0, 0);
ethd->ithread = ithread;
ethd->ptr1 = (void *)cpi;
ethd->ptr2 = (void *)&cpi->mb_row_ei[ithread];
rc = pthread_create(&cpi->h_encoding_thread[ithread], 0,
thread_encoding_proc, ethd);
if(rc)
break;
}
if(rc)
{
/* shutdown other threads */
cpi->b_multi_threaded = 0;
for(--ithread; ithread >= 0; ithread--)
{
pthread_join(cpi->h_encoding_thread[ithread], 0);
sem_destroy(&cpi->h_event_start_encoding[ithread]);
}
sem_destroy(&cpi->h_event_end_encoding);
/* free thread related resources */
vpx_free(cpi->h_event_start_encoding);
vpx_free(cpi->h_encoding_thread);
vpx_free(cpi->mb_row_ei);
vpx_free(cpi->en_thread_data);
return -1;
}
{
LPFTHREAD_DATA * lpfthd = &cpi->lpf_thread_data;
sem_init(&cpi->h_event_start_lpf, 0, 0);
sem_init(&cpi->h_event_end_lpf, 0, 0);
lpfthd->ptr1 = (void *)cpi;
rc = pthread_create(&cpi->h_filter_thread, 0, thread_loopfilter,
lpfthd);
if(rc)
{
/* shutdown other threads */
cpi->b_multi_threaded = 0;
for(--ithread; ithread >= 0; ithread--)
{
sem_post(&cpi->h_event_start_encoding[ithread]);
pthread_join(cpi->h_encoding_thread[ithread], 0);
sem_destroy(&cpi->h_event_start_encoding[ithread]);
}
sem_destroy(&cpi->h_event_end_encoding);
sem_destroy(&cpi->h_event_end_lpf);
sem_destroy(&cpi->h_event_start_lpf);
/* free thread related resources */
vpx_free(cpi->h_event_start_encoding);
vpx_free(cpi->h_encoding_thread);
vpx_free(cpi->mb_row_ei);
vpx_free(cpi->en_thread_data);
return -2;
}
}
}
return 0;
}
void vp9_init_layer_context(VP9_COMP *const cpi) {
SVC *const svc = &cpi->svc;
const VP9EncoderConfig *const oxcf = &cpi->oxcf;
int mi_rows = cpi->common.mi_rows;
int mi_cols = cpi->common.mi_cols;
int sl, tl, i;
int alt_ref_idx = svc->number_spatial_layers;
svc->spatial_layer_id = 0;
svc->temporal_layer_id = 0;
svc->first_spatial_layer_to_encode = 0;
svc->rc_drop_superframe = 0;
svc->force_zero_mode_spatial_ref = 0;
svc->use_base_mv = 0;
svc->scaled_temp_is_alloc = 0;
svc->scaled_one_half = 0;
svc->current_superframe = 0;
for (i = 0; i < REF_FRAMES; ++i)
svc->ref_frame_index[i] = -1;
for (sl = 0; sl < oxcf->ss_number_layers; ++sl) {
cpi->svc.ext_frame_flags[sl] = 0;
cpi->svc.ext_lst_fb_idx[sl] = 0;
cpi->svc.ext_gld_fb_idx[sl] = 1;
cpi->svc.ext_alt_fb_idx[sl] = 2;
}
if (cpi->oxcf.error_resilient_mode == 0 && cpi->oxcf.pass == 2) {
if (vpx_realloc_frame_buffer(&cpi->svc.empty_frame.img,
SMALL_FRAME_WIDTH, SMALL_FRAME_HEIGHT,
cpi->common.subsampling_x,
cpi->common.subsampling_y,
#if CONFIG_VP9_HIGHBITDEPTH
cpi->common.use_highbitdepth,
#endif
VP9_ENC_BORDER_IN_PIXELS,
cpi->common.byte_alignment,
NULL, NULL, NULL))
vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
"Failed to allocate empty frame for multiple frame "
"contexts");
memset(cpi->svc.empty_frame.img.buffer_alloc, 0x80,
cpi->svc.empty_frame.img.buffer_alloc_sz);
}
for (sl = 0; sl < oxcf->ss_number_layers; ++sl) {
for (tl = 0; tl < oxcf->ts_number_layers; ++tl) {
int layer = LAYER_IDS_TO_IDX(sl, tl, oxcf->ts_number_layers);
LAYER_CONTEXT *const lc = &svc->layer_context[layer];
RATE_CONTROL *const lrc = &lc->rc;
int i;
lc->current_video_frame_in_layer = 0;
lc->layer_size = 0;
lc->frames_from_key_frame = 0;
lc->last_frame_type = FRAME_TYPES;
lrc->ni_av_qi = oxcf->worst_allowed_q;
lrc->total_actual_bits = 0;
lrc->total_target_vs_actual = 0;
lrc->ni_tot_qi = 0;
lrc->tot_q = 0.0;
lrc->avg_q = 0.0;
lrc->ni_frames = 0;
lrc->decimation_count = 0;
lrc->decimation_factor = 0;
for (i = 0; i < RATE_FACTOR_LEVELS; ++i) {
lrc->rate_correction_factors[i] = 1.0;
}
if (cpi->oxcf.rc_mode == VPX_CBR) {
lc->target_bandwidth = oxcf->layer_target_bitrate[layer];
lrc->last_q[INTER_FRAME] = oxcf->worst_allowed_q;
lrc->avg_frame_qindex[INTER_FRAME] = oxcf->worst_allowed_q;
lrc->avg_frame_qindex[KEY_FRAME] = oxcf->worst_allowed_q;
} else {
lc->target_bandwidth = oxcf->layer_target_bitrate[layer];
lrc->last_q[KEY_FRAME] = oxcf->best_allowed_q;
lrc->last_q[INTER_FRAME] = oxcf->best_allowed_q;
lrc->avg_frame_qindex[KEY_FRAME] = (oxcf->worst_allowed_q +
oxcf->best_allowed_q) / 2;
lrc->avg_frame_qindex[INTER_FRAME] = (oxcf->worst_allowed_q +
oxcf->best_allowed_q) / 2;
if (oxcf->ss_enable_auto_arf[sl])
lc->alt_ref_idx = alt_ref_idx++;
else
lc->alt_ref_idx = INVALID_IDX;
lc->gold_ref_idx = INVALID_IDX;
}
lrc->buffer_level = oxcf->starting_buffer_level_ms *
lc->target_bandwidth / 1000;
lrc->bits_off_target = lrc->buffer_level;
// Initialize the cyclic refresh parameters. If spatial layers are used
// (i.e., ss_number_layers > 1), these need to be updated per spatial
// layer.
// Cyclic refresh is only applied on base temporal layer.
if (oxcf->ss_number_layers > 1 &&
tl == 0) {
size_t last_coded_q_map_size;
size_t consec_zero_mv_size;
VP9_COMMON *const cm = &cpi->common;
lc->sb_index = 0;
CHECK_MEM_ERROR(cm, lc->map,
vpx_malloc(mi_rows * mi_cols * sizeof(*lc->map)));
memset(lc->map, 0, mi_rows * mi_cols);
last_coded_q_map_size = mi_rows * mi_cols *
sizeof(*lc->last_coded_q_map);
CHECK_MEM_ERROR(cm, lc->last_coded_q_map,
vpx_malloc(last_coded_q_map_size));
assert(MAXQ <= 255);
memset(lc->last_coded_q_map, MAXQ, last_coded_q_map_size);
consec_zero_mv_size = mi_rows * mi_cols * sizeof(*lc->consec_zero_mv);
CHECK_MEM_ERROR(cm, lc->consec_zero_mv,
vpx_malloc(consec_zero_mv_size));
memset(lc->consec_zero_mv, 0, consec_zero_mv_size);
}
}
}
// Still have extra buffer for base layer golden frame
if (!(svc->number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR)
&& alt_ref_idx < REF_FRAMES)
svc->layer_context[0].gold_ref_idx = alt_ref_idx;
}
static void setup_token_decoder(VP8D_COMP *pbi,
const unsigned char* token_part_sizes)
{
vp8_reader *bool_decoder = &pbi->bc2;
unsigned int partition_idx;
int fragment_idx;
int num_token_partitions;
const unsigned char *first_fragment_end = pbi->fragments[0] +
pbi->fragment_sizes[0];
TOKEN_PARTITION multi_token_partition =
(TOKEN_PARTITION)vp8_read_literal(&pbi->bc, 2);
if (!vp8dx_bool_error(&pbi->bc))
pbi->common.multi_token_partition = multi_token_partition;
num_token_partitions = 1 << pbi->common.multi_token_partition;
if (num_token_partitions > 1)
{
CHECK_MEM_ERROR(pbi->mbc, vpx_malloc(num_token_partitions *
sizeof(vp8_reader)));
bool_decoder = pbi->mbc;
}
/* Check for partitions within the fragments and unpack the fragments
* so that each fragment pointer points to its corresponding partition. */
for (fragment_idx = 0; fragment_idx < pbi->num_fragments; ++fragment_idx)
{
unsigned int fragment_size = pbi->fragment_sizes[fragment_idx];
const unsigned char *fragment_end = pbi->fragments[fragment_idx] +
fragment_size;
/* Special case for handling the first partition since we have already
* read its size. */
if (fragment_idx == 0)
{
/* Size of first partition + token partition sizes element */
ptrdiff_t ext_first_part_size = token_part_sizes -
pbi->fragments[0] + 3 * (num_token_partitions - 1);
fragment_size -= ext_first_part_size;
if (fragment_size > 0)
{
pbi->fragment_sizes[0] = ext_first_part_size;
/* The fragment contains an additional partition. Move to
* next. */
fragment_idx++;
pbi->fragments[fragment_idx] = pbi->fragments[0] +
pbi->fragment_sizes[0];
}
}
/* Split the chunk into partitions read from the bitstream */
while (fragment_size > 0)
{
ptrdiff_t partition_size = read_available_partition_size(
pbi,
token_part_sizes,
pbi->fragments[fragment_idx],
first_fragment_end,
fragment_end,
fragment_idx - 1,
num_token_partitions);
pbi->fragment_sizes[fragment_idx] = partition_size;
fragment_size -= partition_size;
assert(fragment_idx <= num_token_partitions);
if (fragment_size > 0)
{
/* The fragment contains an additional partition.
* Move to next. */
fragment_idx++;
pbi->fragments[fragment_idx] =
pbi->fragments[fragment_idx - 1] + partition_size;
}
}
}
pbi->num_fragments = num_token_partitions + 1;
for (partition_idx = 1; partition_idx < pbi->num_fragments; ++partition_idx)
{
if (vp8dx_start_decode(bool_decoder,
pbi->fragments[partition_idx],
pbi->fragment_sizes[partition_idx]))
vpx_internal_error(&pbi->common.error, VPX_CODEC_MEM_ERROR,
"Failed to allocate bool decoder %d",
partition_idx);
bool_decoder++;
}
#if CONFIG_MULTITHREAD
/* Clamp number of decoder threads */
if (pbi->decoding_thread_count > num_token_partitions - 1)
pbi->decoding_thread_count = num_token_partitions - 1;
#endif
}
void vp10_encode_tiles_mt(VP10_COMP *cpi) {
VP10_COMMON *const cm = &cpi->common;
const int tile_cols = 1 << cm->log2_tile_cols;
const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
const int num_workers = VPXMIN(cpi->oxcf.max_threads, tile_cols);
int i;
vp10_init_tile_data(cpi);
// Only run once to create threads and allocate thread data.
if (cpi->num_workers == 0) {
int allocated_workers = num_workers;
CHECK_MEM_ERROR(cm, cpi->workers,
vpx_malloc(allocated_workers * sizeof(*cpi->workers)));
CHECK_MEM_ERROR(cm, cpi->tile_thr_data,
vpx_calloc(allocated_workers,
sizeof(*cpi->tile_thr_data)));
for (i = 0; i < allocated_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *thread_data = &cpi->tile_thr_data[i];
++cpi->num_workers;
winterface->init(worker);
if (i < allocated_workers - 1) {
thread_data->cpi = cpi;
// Allocate thread data.
CHECK_MEM_ERROR(cm, thread_data->td,
vpx_memalign(32, sizeof(*thread_data->td)));
vp10_zero(*thread_data->td);
// Set up pc_tree.
thread_data->td->leaf_tree = NULL;
thread_data->td->pc_tree = NULL;
vp10_setup_pc_tree(cm, thread_data->td);
// Allocate frame counters in thread data.
CHECK_MEM_ERROR(cm, thread_data->td->counts,
vpx_calloc(1, sizeof(*thread_data->td->counts)));
// Create threads
if (!winterface->reset(worker))
vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
"Tile encoder thread creation failed");
} else {
// Main thread acts as a worker and uses the thread data in cpi.
thread_data->cpi = cpi;
thread_data->td = &cpi->td;
}
winterface->sync(worker);
}
}
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *thread_data;
worker->hook = (VPxWorkerHook)enc_worker_hook;
worker->data1 = &cpi->tile_thr_data[i];
worker->data2 = NULL;
thread_data = (EncWorkerData*)worker->data1;
// Before encoding a frame, copy the thread data from cpi.
if (thread_data->td != &cpi->td) {
thread_data->td->mb = cpi->td.mb;
thread_data->td->rd_counts = cpi->td.rd_counts;
}
if (thread_data->td->counts != &cpi->common.counts) {
memcpy(thread_data->td->counts, &cpi->common.counts,
sizeof(cpi->common.counts));
}
}
// Encode a frame
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
// Set the starting tile for each thread.
thread_data->start = i;
if (i == cpi->num_workers - 1)
winterface->execute(worker);
else
winterface->launch(worker);
}
// Encoding ends.
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
winterface->sync(worker);
}
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
// Accumulate counters.
if (i < cpi->num_workers - 1) {
vp10_accumulate_frame_counts(cm, thread_data->td->counts, 0);
accumulate_rd_opt(&cpi->td, thread_data->td);
}
}
}
void vp10_dering_frame(YV12_BUFFER_CONFIG *frame, VP10_COMMON *cm,
MACROBLOCKD *xd, int global_level) {
int r, c;
int sbr, sbc;
int nhsb, nvsb;
od_dering_in *src[3];
unsigned char *bskip;
int dir[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS] = {{0}};
int stride;
int bsize[3];
int dec[3];
int pli;
int coeff_shift = VPXMAX(cm->bit_depth - 8, 0);
nvsb = (cm->mi_rows + MI_BLOCK_SIZE - 1)/MI_BLOCK_SIZE;
nhsb = (cm->mi_cols + MI_BLOCK_SIZE - 1)/MI_BLOCK_SIZE;
bskip = vpx_malloc(sizeof(*bskip)*cm->mi_rows*cm->mi_cols);
vp10_setup_dst_planes(xd->plane, frame, 0, 0);
for (pli = 0; pli < 3; pli++) {
dec[pli] = xd->plane[pli].subsampling_x;
bsize[pli] = 8 >> dec[pli];
}
stride = bsize[0]*cm->mi_cols;
for (pli = 0; pli < 3; pli++) {
src[pli] = vpx_malloc(sizeof(*src)*cm->mi_rows*cm->mi_cols*64);
for (r = 0; r < bsize[pli]*cm->mi_rows; ++r) {
for (c = 0; c < bsize[pli]*cm->mi_cols; ++c) {
#if CONFIG_VPX_HIGHBITDEPTH
if (cm->use_highbitdepth) {
src[pli][r * stride + c] =
CONVERT_TO_SHORTPTR(xd->plane[pli].dst.buf)
[r * xd->plane[pli].dst.stride + c];
} else {
#endif
src[pli][r * stride + c] =
xd->plane[pli].dst.buf[r * xd->plane[pli].dst.stride + c];
#if CONFIG_VPX_HIGHBITDEPTH
}
#endif
}
}
}
for (r = 0; r < cm->mi_rows; ++r) {
for (c = 0; c < cm->mi_cols; ++c) {
const MB_MODE_INFO *mbmi =
&cm->mi_grid_visible[r * cm->mi_stride + c]->mbmi;
bskip[r * cm->mi_cols + c] = mbmi->skip;
}
}
for (sbr = 0; sbr < nvsb; sbr++) {
for (sbc = 0; sbc < nhsb; sbc++) {
int level;
int nhb, nvb;
nhb = VPXMIN(MI_BLOCK_SIZE, cm->mi_cols - MI_BLOCK_SIZE*sbc);
nvb = VPXMIN(MI_BLOCK_SIZE, cm->mi_rows - MI_BLOCK_SIZE*sbr);
for (pli = 0; pli < 3; pli++) {
int16_t dst[MI_BLOCK_SIZE*MI_BLOCK_SIZE*8*8];
int threshold;
#if DERING_REFINEMENT
level = compute_level_from_index(
global_level,
cm->mi_grid_visible[MI_BLOCK_SIZE*sbr*cm->mi_stride +
MI_BLOCK_SIZE*sbc]->mbmi.dering_gain);
#else
level = global_level;
#endif
/* FIXME: This is a temporary hack that uses more conservative
deringing for chroma. */
if (pli) level = (level*5 + 4) >> 3;
if (sb_all_skip(cm, sbr*MI_BLOCK_SIZE, sbc*MI_BLOCK_SIZE)) level = 0;
threshold = level << coeff_shift;
od_dering(
&OD_DERING_VTBL_C,
dst,
MI_BLOCK_SIZE*bsize[pli],
&src[pli][sbr*stride*bsize[pli]*MI_BLOCK_SIZE +
sbc*bsize[pli]*MI_BLOCK_SIZE],
stride, nhb, nvb, sbc, sbr, nhsb, nvsb, dec[pli], dir, pli,
&bskip[MI_BLOCK_SIZE*sbr*cm->mi_cols + MI_BLOCK_SIZE*sbc],
cm->mi_cols, threshold, OD_DERING_NO_CHECK_OVERLAP, coeff_shift);
for (r = 0; r < bsize[pli]*nvb; ++r) {
for (c = 0; c < bsize[pli]*nhb; ++c) {
#if CONFIG_VPX_HIGHBITDEPTH
if (cm->use_highbitdepth) {
CONVERT_TO_SHORTPTR(xd->plane[pli].dst.buf)
[xd->plane[pli].dst.stride*(bsize[pli]*MI_BLOCK_SIZE*sbr + r)
+ sbc*bsize[pli]*MI_BLOCK_SIZE + c] =
dst[r * MI_BLOCK_SIZE * bsize[pli] + c];
} else {
#endif
xd->plane[pli].dst.buf[xd->plane[pli].dst.stride*
(bsize[pli]*MI_BLOCK_SIZE*sbr + r) +
sbc*bsize[pli]*MI_BLOCK_SIZE + c] =
dst[r * MI_BLOCK_SIZE * bsize[pli] + c];
#if CONFIG_VPX_HIGHBITDEPTH
}
#endif
}
}
}
}
}
for (pli = 0; pli < 3; pli++) {
vpx_free(src[pli]);
}
vpx_free(bskip);
}
void vp8cx_create_encoder_threads(VP8_COMP *cpi)
{
const VP8_COMMON * cm = &cpi->common;
cpi->b_multi_threaded = 0;
cpi->encoding_thread_count = 0;
if (cm->processor_core_count > 1 && cpi->oxcf.multi_threaded > 1)
{
int ithread;
int th_count = cpi->oxcf.multi_threaded - 1;
/* don't allocate more threads than cores available */
if (cpi->oxcf.multi_threaded > cm->processor_core_count)
th_count = cm->processor_core_count - 1;
/* we have th_count + 1 (main) threads processing one row each */
/* no point to have more threads than the sync range allows */
if(th_count > ((cm->mb_cols / cpi->mt_sync_range) - 1))
{
th_count = (cm->mb_cols / cpi->mt_sync_range) - 1;
}
if(th_count == 0)
return;
CHECK_MEM_ERROR(cpi->h_encoding_thread, vpx_malloc(sizeof(pthread_t) * th_count));
CHECK_MEM_ERROR(cpi->h_event_start_encoding, vpx_malloc(sizeof(sem_t) * th_count));
CHECK_MEM_ERROR(cpi->mb_row_ei, vpx_memalign(32, sizeof(MB_ROW_COMP) * th_count));
vpx_memset(cpi->mb_row_ei, 0, sizeof(MB_ROW_COMP) * th_count);
CHECK_MEM_ERROR(cpi->en_thread_data,
vpx_malloc(sizeof(ENCODETHREAD_DATA) * th_count));
CHECK_MEM_ERROR(cpi->mt_current_mb_col,
vpx_malloc(sizeof(*cpi->mt_current_mb_col) * cm->mb_rows));
sem_init(&cpi->h_event_end_encoding, 0, 0);
cpi->b_multi_threaded = 1;
cpi->encoding_thread_count = th_count;
/*
printf("[VP8:] multi_threaded encoding is enabled with %d threads\n\n",
(cpi->encoding_thread_count +1));
*/
for (ithread = 0; ithread < th_count; ithread++)
{
ENCODETHREAD_DATA * ethd = &cpi->en_thread_data[ithread];
sem_init(&cpi->h_event_start_encoding[ithread], 0, 0);
ethd->ithread = ithread;
ethd->ptr1 = (void *)cpi;
ethd->ptr2 = (void *)&cpi->mb_row_ei[ithread];
pthread_create(&cpi->h_encoding_thread[ithread], 0, thread_encoding_proc, ethd);
}
{
LPFTHREAD_DATA * lpfthd = &cpi->lpf_thread_data;
sem_init(&cpi->h_event_start_lpf, 0, 0);
sem_init(&cpi->h_event_end_lpf, 0, 0);
lpfthd->ptr1 = (void *)cpi;
pthread_create(&cpi->h_filter_thread, 0, loopfilter_thread, lpfthd);
}
}
}
void vp9_encode_tiles_mt(VP9_COMP *cpi) {
VP9_COMMON *const cm = &cpi->common;
const int tile_cols = 1 << cm->log2_tile_cols;
const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
const int num_workers = VPXMIN(cpi->oxcf.max_threads, tile_cols);
int i;
vp9_init_tile_data(cpi);
// Only run once to create threads and allocate thread data.
if (cpi->num_workers == 0) {
int allocated_workers = num_workers;
// While using SVC, we need to allocate threads according to the highest
// resolution.
if (cpi->use_svc) {
int max_tile_cols = get_max_tile_cols(cpi);
allocated_workers = VPXMIN(cpi->oxcf.max_threads, max_tile_cols);
}
CHECK_MEM_ERROR(cm, cpi->workers,
vpx_malloc(allocated_workers * sizeof(*cpi->workers)));
CHECK_MEM_ERROR(cm, cpi->tile_thr_data,
vpx_calloc(allocated_workers,
sizeof(*cpi->tile_thr_data)));
for (i = 0; i < allocated_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *thread_data = &cpi->tile_thr_data[i];
++cpi->num_workers;
winterface->init(worker);
if (i < allocated_workers - 1) {
thread_data->cpi = cpi;
// Allocate thread data.
CHECK_MEM_ERROR(cm, thread_data->td,
vpx_memalign(32, sizeof(*thread_data->td)));
vp9_zero(*thread_data->td);
// Set up pc_tree.
thread_data->td->leaf_tree = NULL;
thread_data->td->pc_tree = NULL;
vp9_setup_pc_tree(cm, thread_data->td);
// Allocate frame counters in thread data.
CHECK_MEM_ERROR(cm, thread_data->td->counts,
vpx_calloc(1, sizeof(*thread_data->td->counts)));
// Create threads
if (!winterface->reset(worker))
vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
"Tile encoder thread creation failed");
} else {
// Main thread acts as a worker and uses the thread data in cpi.
thread_data->cpi = cpi;
thread_data->td = &cpi->td;
}
winterface->sync(worker);
}
}
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *thread_data;
worker->hook = (VPxWorkerHook)enc_worker_hook;
worker->data1 = &cpi->tile_thr_data[i];
worker->data2 = NULL;
thread_data = (EncWorkerData*)worker->data1;
// Before encoding a frame, copy the thread data from cpi.
if (thread_data->td != &cpi->td) {
thread_data->td->mb = cpi->td.mb;
thread_data->td->rd_counts = cpi->td.rd_counts;
}
if (thread_data->td->counts != &cpi->common.counts) {
memcpy(thread_data->td->counts, &cpi->common.counts,
sizeof(cpi->common.counts));
}
// Handle use_nonrd_pick_mode case.
if (cpi->sf.use_nonrd_pick_mode) {
MACROBLOCK *const x = &thread_data->td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
struct macroblock_plane *const p = x->plane;
struct macroblockd_plane *const pd = xd->plane;
PICK_MODE_CONTEXT *ctx = &thread_data->td->pc_root->none;
int j;
for (j = 0; j < MAX_MB_PLANE; ++j) {
p[j].coeff = ctx->coeff_pbuf[j][0];
p[j].qcoeff = ctx->qcoeff_pbuf[j][0];
pd[j].dqcoeff = ctx->dqcoeff_pbuf[j][0];
p[j].eobs = ctx->eobs_pbuf[j][0];
}
}
}
// Encode a frame
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
// Set the starting tile for each thread.
thread_data->start = i;
if (i == cpi->num_workers - 1)
winterface->execute(worker);
else
winterface->launch(worker);
}
// Encoding ends.
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
winterface->sync(worker);
}
for (i = 0; i < num_workers; i++) {
VPxWorker *const worker = &cpi->workers[i];
EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
// Accumulate counters.
if (i < cpi->num_workers - 1) {
vp9_accumulate_frame_counts(cm, thread_data->td->counts, 0);
accumulate_rd_opt(&cpi->td, thread_data->td);
}
}
}
