4 files changed, 36 insertions, 16 deletions

diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index 7cfe0de90..35fea57f5 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -1786,7 +1786,7 @@ void vp9_remove_compressor(VP9_COMP *cpi) {
 
   for (t = 0; t < cpi->num_workers; ++t) {
     VP9Worker *const worker = &cpi->workers[t];
-    EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
+    EncWorkerData *const thread_data = &cpi->tile_thr_data[t];
 
     // Deallocate allocated threads.
     vp9_get_worker_interface()->end(worker);
@@ -1797,11 +1797,13 @@ void vp9_remove_compressor(VP9_COMP *cpi) {
       vp9_free_pc_tree(thread_data->td);
       vpx_free(thread_data->td);
     }
-
-    vpx_free(worker->data1);
   }
+  vpx_free(cpi->tile_thr_data);
   vpx_free(cpi->workers);
 
+  if (cpi->num_workers > 1)
+    vp9_loop_filter_dealloc(&cpi->lf_row_sync);
+
   dealloc_compressor_data(cpi);
 
   for (i = 0; i < sizeof(cpi->mbgraph_stats) /
@@ -2437,7 +2439,13 @@ static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) {
   }
 
   if (lf->filter_level > 0) {
-    vp9_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
+    if (cpi->num_workers > 1)
+      vp9_loop_filter_frame_mt(cm->frame_to_show, cm, xd->plane,
+                               lf->filter_level, 0, 0,
+                               cpi->workers, cpi->num_workers,
+                               &cpi->lf_row_sync);
+    else
+      vp9_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
   }
 
   vp9_extend_frame_inner_borders(cm->frame_to_show);
diff --git a/vp9/encoder/vp9_encoder.h b/vp9/encoder/vp9_encoder.h
index a41284af0..cf269c108 100644
--- a/vp9/encoder/vp9_encoder.h
+++ b/vp9/encoder/vp9_encoder.h
@@ -19,6 +19,7 @@
 
 #include "vp9/common/vp9_ppflags.h"
 #include "vp9/common/vp9_entropymode.h"
+#include "vp9/common/vp9_loopfilter_thread.h"
 #include "vp9/common/vp9_onyxc_int.h"
 #include "vp9/common/vp9_thread.h"
 
@@ -36,6 +37,7 @@
 #include "vp9/encoder/vp9_svc_layercontext.h"
 #include "vp9/encoder/vp9_tokenize.h"
 #include "vp9/encoder/vp9_variance.h"
+
 #if CONFIG_VP9_TEMPORAL_DENOISING
 #include "vp9/encoder/vp9_denoiser.h"
 #endif
@@ -262,6 +264,8 @@ typedef struct ThreadData {
   PC_TREE *pc_root;
 } ThreadData;
 
+struct EncWorkerData;
+
 typedef struct VP9_COMP {
   QUANTS quants;
   ThreadData td;
@@ -447,6 +451,8 @@ typedef struct VP9_COMP {
   // Multi-threading
   int num_workers;
   VP9Worker *workers;
+  struct EncWorkerData *tile_thr_data;
+  VP9LfSync lf_row_sync;
 } VP9_COMP;
 
 void vp9_initialize_enc(void);
diff --git a/vp9/encoder/vp9_ethread.c b/vp9/encoder/vp9_ethread.c
index 9ee32366c..12fb4d107 100644
--- a/vp9/encoder/vp9_ethread.c
+++ b/vp9/encoder/vp9_ethread.c
@@ -167,16 +167,15 @@ void vp9_encode_tiles_mt(VP9_COMP *cpi) {
     CHECK_MEM_ERROR(cm, cpi->workers,
                     vpx_malloc(num_workers * sizeof(*cpi->workers)));
 
+    CHECK_MEM_ERROR(cm, cpi->tile_thr_data,
+                    vpx_calloc(num_workers, sizeof(*cpi->tile_thr_data)));
+
     for (i = 0; i < num_workers; i++) {
       VP9Worker *const worker = &cpi->workers[i];
-      EncWorkerData *thread_data;
+      EncWorkerData *thread_data = &cpi->tile_thr_data[i];
 
       ++cpi->num_workers;
-
       winterface->init(worker);
-      CHECK_MEM_ERROR(cm, worker->data1,
-                      (EncWorkerData*)vpx_calloc(1, sizeof(EncWorkerData)));
-      thread_data = (EncWorkerData*)worker->data1;
 
       if (i < num_workers - 1) {
       thread_data->cpi = cpi;
@@ -205,17 +204,18 @@ void vp9_encode_tiles_mt(VP9_COMP *cpi) {
         thread_data->td = &cpi->td;
       }
 
-      // data2 is unused.
-      worker->data2 = NULL;
-
       winterface->sync(worker);
-      worker->hook = (VP9WorkerHook)enc_worker_hook;
     }
   }
 
   for (i = 0; i < num_workers; i++) {
     VP9Worker *const worker = &cpi->workers[i];
-    EncWorkerData *const thread_data = (EncWorkerData*)worker->data1;
+    EncWorkerData *thread_data;
+
+    worker->hook = (VP9WorkerHook)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.
     thread_data->td->mb = cpi->td.mb;
diff --git a/vp9/encoder/vp9_picklpf.c b/vp9/encoder/vp9_picklpf.c
index 01cc519f9..a95f0f46d 100644
--- a/vp9/encoder/vp9_picklpf.c
+++ b/vp9/encoder/vp9_picklpf.c
@@ -39,8 +39,14 @@ static int64_t try_filter_frame(const YV12_BUFFER_CONFIG *sd,
   VP9_COMMON *const cm = &cpi->common;
   int64_t filt_err;
 
-  vp9_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filt_level, 1,
-                        partial_frame);
+  if (cpi->num_workers > 1)
+    vp9_loop_filter_frame_mt(cm->frame_to_show, cm, cpi->td.mb.e_mbd.plane,
+                             filt_level, 1, partial_frame,
+                             cpi->workers, cpi->num_workers, &cpi->lf_row_sync);
+  else
+    vp9_loop_filter_frame(cm->frame_to_show, cm, &cpi->td.mb.e_mbd, filt_level,
+                          1, partial_frame);
+
 #if CONFIG_VP9_HIGHBITDEPTH
   if (cm->use_highbitdepth) {
     filt_err = vp9_highbd_get_y_sse(sd, cm->frame_to_show);