vp9/decoder: threaded row-based loop filter

Currently the only threaded option for vp9 decode. Enabled when the
decoder config thread count is > 1.

Change-Id: I082959abac9e31aa4a38ed9fd68b94680e57f4df

6 files changed, 93 insertions, 4 deletions

diff --git a/test/vp9_thread_test.cc b/test/vp9_thread_test.cc
index 308ba9438..41d22dd3a 100644
--- a/test/vp9_thread_test.cc
+++ b/test/vp9_thread_test.cc
@@ -11,6 +11,10 @@
 #include "vp9/decoder/vp9_thread.h"
 
 #include "third_party/googletest/src/include/gtest/gtest.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/md5_helper.h"
+#include "test/webm_video_source.h"
 
 namespace {
 
@@ -77,4 +81,29 @@ TEST_F(VP9WorkerThreadTest, HookFailure) {
   EXPECT_FALSE(worker_.had_error);
 }
 
+TEST(VP9DecodeMTTest, MTDecode) {
+  libvpx_test::WebMVideoSource video("vp90-2-03-size-226x226.webm");
+  video.Init();
+
+  vpx_codec_dec_cfg_t cfg = {0};
+  cfg.threads = 2;
+  libvpx_test::VP9Decoder decoder(cfg, 0);
+
+  libvpx_test::MD5 md5;
+  for (video.Begin(); video.cxdata(); video.Next()) {
+    const vpx_codec_err_t res =
+        decoder.DecodeFrame(video.cxdata(), video.frame_size());
+    ASSERT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+
+    libvpx_test::DxDataIterator dec_iter = decoder.GetDxData();
+    const vpx_image_t *img = NULL;
+
+    // Get decompressed data
+    while ((img = dec_iter.Next())) {
+      md5.Add(img);
+    }
+  }
+  EXPECT_STREQ("b35a1b707b28e82be025d960aba039bc", md5.Get());
+}
+
 }  // namespace
diff --git a/vp9/common/vp9_loopfilter.c b/vp9/common/vp9_loopfilter.c
index 5498b1717..c57f0a55d 100644
--- a/vp9/common/vp9_loopfilter.c
+++ b/vp9/common/vp9_loopfilter.c
@@ -376,3 +376,11 @@ void vp9_loop_filter_frame(VP9_COMMON *cm, MACROBLOCKD *xd,
   vp9_loop_filter_rows(cm->frame_to_show, cm, xd,
                        0, cm->mi_rows, y_only);
 }
+
+int vp9_loop_filter_worker(void *arg1, void *arg2) {
+  LFWorkerData *const lf_data = (LFWorkerData*)arg1;
+  (void)arg2;
+  vp9_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, &lf_data->xd,
+                       lf_data->start, lf_data->stop, lf_data->y_only);
+  return 1;
+}
diff --git a/vp9/common/vp9_loopfilter.h b/vp9/common/vp9_loopfilter.h
index e59cc6485..c6fe112ec 100644
--- a/vp9/common/vp9_loopfilter.h
+++ b/vp9/common/vp9_loopfilter.h
@@ -64,4 +64,18 @@ void vp9_loop_filter_frame(struct VP9Common *cm,
 void vp9_loop_filter_rows(const YV12_BUFFER_CONFIG *frame_buffer,
                           struct VP9Common *cm, struct macroblockd *xd,
                           int start, int stop, int y_only);
+
+typedef struct LoopFilterWorkerData {
+  const YV12_BUFFER_CONFIG *frame_buffer;
+  struct VP9Common *cm;
+  struct macroblockd xd;  // TODO(jzern): most of this is unnecessary to the
+                          // loopfilter. the planes are necessary as their state
+                          // is changed during decode.
+  int start;
+  int stop;
+  int y_only;
+} LFWorkerData;
+
+// Operates on the rows described by LFWorkerData passed as 'arg1'.
+int vp9_loop_filter_worker(void *arg1, void *arg2);
 #endif  // VP9_COMMON_VP9_LOOPFILTER_H_
diff --git a/vp9/decoder/vp9_decodframe.c b/vp9/decoder/vp9_decodframe.c
index ff7cb8d34..2b6f5a9c6 100644
--- a/vp9/decoder/vp9_decodframe.c
+++ b/vp9/decoder/vp9_decodframe.c
@@ -34,6 +34,7 @@
 #include "vp9/decoder/vp9_idct_blk.h"
 #include "vp9/decoder/vp9_onyxd_int.h"
 #include "vp9/decoder/vp9_read_bit_buffer.h"
+#include "vp9/decoder/vp9_thread.h"
 #include "vp9/decoder/vp9_treereader.h"
 
 static int read_be32(const uint8_t *p) {
@@ -585,10 +586,18 @@ static void setup_frame_size_with_refs(VP9D_COMP *pbi,
 }
 
 static void decode_tile(VP9D_COMP *pbi, vp9_reader *r) {
+  const int num_threads = pbi->oxcf.max_threads;
   VP9_COMMON *const pc = &pbi->common;
   int mi_row, mi_col;
 
   if (pbi->do_loopfilter_inline) {
+    if (num_threads > 1) {
+      LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+      lf_data->frame_buffer = &pbi->common.yv12_fb[pbi->common.new_fb_idx];
+      lf_data->cm = pc;
+      lf_data->xd = pbi->mb;
+      lf_data->y_only = 0;
+    }
     vp9_loop_filter_frame_init(pc, &pbi->mb, pbi->mb.lf.filter_level);
   }
 
@@ -603,17 +612,33 @@ static void decode_tile(VP9D_COMP *pbi, vp9_reader *r) {
     }
 
     if (pbi->do_loopfilter_inline) {
-      YV12_BUFFER_CONFIG *const fb =
-          &pbi->common.yv12_fb[pbi->common.new_fb_idx];
       // delay the loopfilter by 1 macroblock row.
       const int lf_start = mi_row - MI_BLOCK_SIZE;
       if (lf_start < 0) continue;
-      vp9_loop_filter_rows(fb, pc, &pbi->mb, lf_start, mi_row, 0);
+
+      if (num_threads > 1) {
+        LFWorkerData *const lf_data = (LFWorkerData*)pbi->lf_worker.data1;
+
+        vp9_worker_sync(&pbi->lf_worker);
+        lf_data->start = lf_start;
+        lf_data->stop = mi_row;
+        pbi->lf_worker.hook = vp9_loop_filter_worker;
+        vp9_worker_launch(&pbi->lf_worker);
+      } else {
+        YV12_BUFFER_CONFIG *const fb =
+            &pbi->common.yv12_fb[pbi->common.new_fb_idx];
+        vp9_loop_filter_rows(fb, pc, &pbi->mb, lf_start, mi_row, 0);
+      }
     }
   }
 
   if (pbi->do_loopfilter_inline) {
     YV12_BUFFER_CONFIG *const fb = &pbi->common.yv12_fb[pbi->common.new_fb_idx];
+    if (num_threads > 1) {
+      // TODO(jzern): since the loop filter is delayed one mb row, this will be
+      // forced to wait for the last row scheduled in the for loop.
+      vp9_worker_sync(&pbi->lf_worker);
+    }
     vp9_loop_filter_rows(fb, pc, &pbi->mb,
                          mi_row - MI_BLOCK_SIZE, pc->mi_rows, 0);
   }
diff --git a/vp9/decoder/vp9_onyxd_if.c b/vp9/decoder/vp9_onyxd_if.c
index aefb56f9a..5a01dd790 100644
--- a/vp9/decoder/vp9_onyxd_if.c
+++ b/vp9/decoder/vp9_onyxd_if.c
@@ -141,6 +141,16 @@ VP9D_PTR vp9_create_decompressor(VP9D_CONFIG *oxcf) {
   pbi->common.error.setjmp = 0;
   pbi->decoded_key_frame = 0;
 
+  if (pbi->oxcf.max_threads > 1) {
+    vp9_worker_init(&pbi->lf_worker);
+    pbi->lf_worker.data1 = vpx_malloc(sizeof(LFWorkerData));
+    pbi->lf_worker.hook = (VP9WorkerHook)vp9_loop_filter_worker;
+    if (pbi->lf_worker.data1 == NULL || !vp9_worker_reset(&pbi->lf_worker)) {
+      vp9_remove_decompressor(pbi);
+      return NULL;
+    }
+  }
+
   return pbi;
 }
 
@@ -154,6 +164,8 @@ void vp9_remove_decompressor(VP9D_PTR ptr) {
     vpx_free(pbi->common.last_frame_seg_map);
 
   vp9_remove_common(&pbi->common);
+  vp9_worker_end(&pbi->lf_worker);
+  vpx_free(pbi->lf_worker.data1);
   vpx_free(pbi);
 }
 
diff --git a/vp9/decoder/vp9_onyxd_int.h b/vp9/decoder/vp9_onyxd_int.h
index 607d14c13..a051971a1 100644
--- a/vp9/decoder/vp9_onyxd_int.h
+++ b/vp9/decoder/vp9_onyxd_int.h
@@ -14,8 +14,8 @@
 #include "./vpx_config.h"
 
 #include "vp9/common/vp9_onyxc_int.h"
-
 #include "vp9/decoder/vp9_onyxd.h"
+#include "vp9/decoder/vp9_thread.h"
 
 typedef struct VP9Decompressor {
   DECLARE_ALIGNED(16, MACROBLOCKD, mb);
@@ -38,6 +38,7 @@ typedef struct VP9Decompressor {
   int initial_height;
 
   int do_loopfilter_inline;  // apply loopfilter to available rows immediately
+  VP9Worker lf_worker;
 } VP9D_COMP;
 
 #endif  // VP9_DECODER_VP9_TREEREADER_H_