ref: b106abe570a5dcef13805593873d5c0e1350453b
parent: 259e835b1bb365a0fe179f9f394192261069ec88
parent: 91f01a2060619183a35eb6ed3eb15733bc6d1e18
author: Yunqing Wang <yunqingwang@google.com>
date: Tue Feb 7 14:55:40 EST 2017
Merge "Row based multi-threading of ARNR filtering stage"
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -1581,6 +1581,11 @@
cpi->oxcf.speed < 5 && cpi->oxcf.pass == 1) &&
cpi->oxcf.new_mt && !cpi->use_svc)
cpi->new_mt = 1;
+
+ if (cpi->oxcf.mode == GOOD && cpi->oxcf.speed < 5 &&
+ (cpi->oxcf.pass == 0 || cpi->oxcf.pass == 2) && cpi->oxcf.new_mt &&
+ !cpi->use_svc)
+ cpi->new_mt = 1;
}
#ifndef M_LOG2_E
--- a/vp9/encoder/vp9_ethread.c
+++ b/vp9/encoder/vp9_ethread.c
@@ -13,6 +13,7 @@
#include "vp9/encoder/vp9_ethread.h"
#include "vp9/encoder/vp9_firstpass.h"
#include "vp9/encoder/vp9_multi_thread.h"
+#include "vp9/encoder/vp9_temporal_filter.h"
#include "vpx_dsp/vpx_dsp_common.h"
static void accumulate_rd_opt(ThreadData *td, ThreadData *td_t) {@@ -463,4 +464,81 @@
TileDataEnc *this_tile = &cpi->tile_data[i];
accumulate_fp_tile_stat(first_tile_col, this_tile);
}
+}
+
+static int temporal_filter_worker_hook(EncWorkerData *const thread_data,
+ MultiThreadHandle *multi_thread_ctxt) {+ VP9_COMP *const cpi = thread_data->cpi;
+ const VP9_COMMON *const cm = &cpi->common;
+ const int tile_cols = 1 << cm->log2_tile_cols;
+ int tile_row, tile_col;
+ int mb_col_start, mb_col_end;
+ TileDataEnc *this_tile;
+ int end_of_frame;
+ int thread_id = thread_data->thread_id;
+ int cur_tile_id = multi_thread_ctxt->thread_id_to_tile_id[thread_id];
+ JobNode *proc_job = NULL;
+ int mb_row;
+
+ end_of_frame = 0;
+ while (0 == end_of_frame) {+ // Get the next job in the queue
+ proc_job =
+ (JobNode *)vp9_enc_grp_get_next_job(multi_thread_ctxt, cur_tile_id);
+ if (NULL == proc_job) {+ // Query for the status of other tiles
+ end_of_frame = vp9_get_tiles_proc_status(
+ multi_thread_ctxt, thread_data->tile_completion_status, &cur_tile_id,
+ tile_cols);
+ } else {+ tile_col = proc_job->tile_col_id;
+ tile_row = proc_job->tile_row_id;
+ this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
+ mb_col_start = (this_tile->tile_info.mi_col_start) >> 1;
+ mb_col_end = (this_tile->tile_info.mi_col_end + 1) >> 1;
+ mb_row = proc_job->vert_unit_row_num;
+
+ vp9_temporal_filter_iterate_row_c(cpi, thread_data->td, mb_row,
+ mb_col_start, mb_col_end);
+ }
+ }
+ return 0;
+}
+
+void vp9_temporal_filter_row_mt(VP9_COMP *cpi) {+ VP9_COMMON *const cm = &cpi->common;
+ const int tile_cols = 1 << cm->log2_tile_cols;
+ const int tile_rows = 1 << cm->log2_tile_rows;
+ MultiThreadHandle *multi_thread_ctxt = &cpi->multi_thread_ctxt;
+ int num_workers = cpi->num_workers ? cpi->num_workers : 1;
+ int i;
+
+ if (multi_thread_ctxt->allocated_tile_cols < tile_cols ||
+ multi_thread_ctxt->allocated_tile_rows < tile_rows ||
+ multi_thread_ctxt->allocated_vert_unit_rows < cm->mb_rows) {+ vp9_row_mt_mem_dealloc(cpi);
+ vp9_init_tile_data(cpi);
+ vp9_row_mt_mem_alloc(cpi);
+ } else {+ vp9_init_tile_data(cpi);
+ }
+
+ create_enc_workers(cpi, num_workers);
+
+ vp9_assign_tile_to_thread(multi_thread_ctxt, tile_cols, cpi->num_workers);
+
+ vp9_prepare_job_queue(cpi, ARNR_JOB);
+
+ for (i = 0; i < num_workers; i++) {+ EncWorkerData *thread_data;
+ thread_data = &cpi->tile_thr_data[i];
+
+ // Before encoding a frame, copy the thread data from cpi.
+ if (thread_data->td != &cpi->td) {+ thread_data->td->mb = cpi->td.mb;
+ }
+ }
+
+ launch_enc_workers(cpi, (VPxWorkerHook)temporal_filter_worker_hook,
+ multi_thread_ctxt, num_workers);
}
--- a/vp9/encoder/vp9_ethread.h
+++ b/vp9/encoder/vp9_ethread.h
@@ -61,6 +61,8 @@
// Deallocate row based multi-threading synchronization related mutex and data.
void vp9_row_mt_sync_mem_dealloc(VP9RowMTSync *row_mt_sync);
+void vp9_temporal_filter_row_mt(struct VP9_COMP *cpi);
+
#ifdef __cplusplus
} // extern "C"
#endif
--- a/vp9/encoder/vp9_mcomp.c
+++ b/vp9/encoder/vp9_mcomp.c
@@ -2381,9 +2381,6 @@
cost_list[4] = INT_MAX;
}
- // Keep track of number of searches (this frame in this thread).
- ++(*x->m_search_count_ptr);
-
switch (method) {case FAST_DIAMOND:
var = fast_dia_search(x, mvp_full, step_param, error_per_bit, 0,
@@ -2409,6 +2406,9 @@
var = full_pixel_diamond(cpi, x, mvp_full, step_param, error_per_bit,
MAX_MVSEARCH_STEPS - 1 - step_param, 1,
cost_list, fn_ptr, ref_mv, tmp_mv);
+
+ // Keep track of number of searches (this frame in this thread).
+ ++(*x->m_search_count_ptr);
// Should we allow a follow on exhaustive search?
if (is_exhaustive_allowed(cpi, x)) {--- a/vp9/encoder/vp9_temporal_filter.c
+++ b/vp9/encoder/vp9_temporal_filter.c
@@ -16,6 +16,7 @@
#include "vp9/common/vp9_quant_common.h"
#include "vp9/common/vp9_reconinter.h"
#include "vp9/encoder/vp9_encodeframe.h"
+#include "vp9/encoder/vp9_ethread.h"
#include "vp9/encoder/vp9_extend.h"
#include "vp9/encoder/vp9_firstpass.h"
#include "vp9/encoder/vp9_mcomp.h"
@@ -262,9 +263,9 @@
return bestsme;
}
-static void temporal_filter_iterate_row_c(VP9_COMP *cpi, ThreadData *td,
- int mb_row, int mb_col_start,
- int mb_col_end) {+void vp9_temporal_filter_iterate_row_c(VP9_COMP *cpi, ThreadData *td,
+ int mb_row, int mb_col_start,
+ int mb_col_end) {ARNRFilterData *arnr_filter_data = &cpi->arnr_filter_data;
YV12_BUFFER_CONFIG **frames = arnr_filter_data->frames;
int frame_count = arnr_filter_data->frame_count;
@@ -571,8 +572,8 @@
int mb_row;
for (mb_row = mb_row_start; mb_row < mb_row_end; mb_row++) {- temporal_filter_iterate_row_c(cpi, &cpi->td, mb_row, mb_col_start,
- mb_col_end);
+ vp9_temporal_filter_iterate_row_c(cpi, &cpi->td, mb_row, mb_col_start,
+ mb_col_end);
}
}
@@ -765,5 +766,8 @@
set_error_per_bit(&cpi->td.mb, rdmult);
vp9_initialize_me_consts(cpi, &cpi->td.mb, ARNR_FILT_QINDEX);
- temporal_filter_iterate_c(cpi);
+ if (!cpi->new_mt)
+ temporal_filter_iterate_c(cpi);
+ else
+ vp9_temporal_filter_row_mt(cpi);
}
--- a/vp9/encoder/vp9_temporal_filter.h
+++ b/vp9/encoder/vp9_temporal_filter.h
@@ -20,6 +20,10 @@
void vp9_temporal_filter_init(void);
void vp9_temporal_filter(VP9_COMP *cpi, int distance);
+void vp9_temporal_filter_iterate_row_c(VP9_COMP *cpi, ThreadData *td,
+ int mb_row, int mb_col_start,
+ int mb_col_end);
+
#ifdef __cplusplus
} // extern "C"
#endif