Merge "Fix sub8x8 motion search on scaled reference frame" into sandbox/jingning@google.com/decoder_test_suite

This commit fixes a buffer overflow issue related to sub8x8 motion
search on scaled reference frame.

Change-Id: Iffeebc8787c1ed2b0cb3a7821349e028639d9eb5

configure

@@ -251,6 +251,8 @@ HAVE_LIST="
 EXPERIMENT_LIST="
     spatial_svc
     fp_mb_stats
+    full_buffer_test
+    internal_resize
     emulate_hardware
 "
 CONFIG_LIST="

test/test.mk

@@ -159,6 +159,7 @@ endif
 ifeq ($(CONFIG_VP9_ENCODER)$(CONFIG_VP9_TEMPORAL_DENOISING),yesyes)
 LIBVPX_TEST_SRCS-$(HAVE_SSE2) += vp9_denoiser_sse2_test.cc
 endif
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += vp9_arf_freq_test.cc
 
 endif # VP9
 

test/test_vector_test.cc

@@ -135,6 +135,7 @@ TEST_P(TestVectorTest, MD5Match) {
 
 // Test VP8 decode in serial mode with single thread.
 // NOTE: VP8 only support serial mode.
+#if CONFIG_VP8_DECODER
 VP8_INSTANTIATE_TEST_CASE(
     TestVectorTest,
     ::testing::Combine(
@@ -143,8 +144,10 @@ VP8_INSTANTIATE_TEST_CASE(
         ::testing::ValuesIn(libvpx_test::kVP8TestVectors,
                             libvpx_test::kVP8TestVectors +
                                 libvpx_test::kNumVP8TestVectors)));
+#endif
 
 // Test VP9 decode in serial mode with single thread.
+#if CONFIG_VP9_DECODER
 VP9_INSTANTIATE_TEST_CASE(
     TestVectorTest,
     ::testing::Combine(
@@ -154,8 +157,6 @@ VP9_INSTANTIATE_TEST_CASE(
                             libvpx_test::kVP9TestVectors +
                                 libvpx_test::kNumVP9TestVectors)));
 
-
-#if CONFIG_VP9_DECODER
 // Test VP9 decode in frame parallel mode with different number of threads.
 INSTANTIATE_TEST_CASE_P(
     VP9MultiThreadedFrameParallel, TestVectorTest,

test/vp9_arf_freq_test.cc

@@ -0,0 +1,230 @@
+/*
+ *  Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "test/codec_factory.h"
+#include "test/encode_test_driver.h"
+#include "test/y4m_video_source.h"
+#include "test/yuv_video_source.h"
+#include "test/util.h"
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "vp9/encoder/vp9_ratectrl.h"
+
+namespace {
+
+const unsigned int kFrames = 100;
+const int kBitrate = 500;
+
+#define ARF_NOT_SEEN   1000001
+#define ARF_SEEN_ONCE  1000000
+
+typedef struct {
+  const char *filename;
+  unsigned int width;
+  unsigned int height;
+  unsigned int framerate_num;
+  unsigned int framerate_den;
+  unsigned int input_bit_depth;
+  vpx_img_fmt fmt;
+  vpx_bit_depth_t bit_depth;
+  unsigned int profile;
+} TestVideoParam;
+
+typedef struct {
+  libvpx_test::TestMode mode;
+  int cpu_used;
+} TestEncodeParam;
+
+const TestVideoParam kTestVectors[] = {
+  // artificially increase framerate to trigger default check
+  {"hantro_collage_w352h288.yuv", 352, 288, 5000, 1,
+    8, VPX_IMG_FMT_I420, VPX_BITS_8, 0},
+  {"hantro_collage_w352h288.yuv", 352, 288, 30, 1,
+    8, VPX_IMG_FMT_I420, VPX_BITS_8, 0},
+  {"rush_hour_444.y4m", 352, 288, 30, 1,
+    8, VPX_IMG_FMT_I444, VPX_BITS_8, 1},
+#if CONFIG_VP9_HIGHBITDEPTH
+  // Add list of profile 2/3 test videos here ...
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+};
+
+const TestEncodeParam kEncodeVectors[] = {
+  {::libvpx_test::kOnePassGood, 2},
+  {::libvpx_test::kOnePassGood, 5},
+  {::libvpx_test::kTwoPassGood, 1},
+  {::libvpx_test::kTwoPassGood, 2},
+  {::libvpx_test::kTwoPassGood, 5},
+  {::libvpx_test::kRealTime, 5},
+};
+
+const int kMinArfVectors[] = {
+  // NOTE: 0 refers to the default built-in logic in:
+  //       vp9_rc_get_default_min_gf_interval(...)
+  0, 4, 8, 12, 15
+};
+
+int is_extension_y4m(const char *filename) {
+  const char *dot = strrchr(filename, '.');
+  if (!dot || dot == filename)
+    return 0;
+  else
+    return !strcmp(dot, ".y4m");
+}
+
+class ArfFreqTest
+    : public ::libvpx_test::EncoderTest,
+      public ::libvpx_test::CodecTestWith3Params<TestVideoParam, \
+                                                 TestEncodeParam, int> {
+ protected:
+  ArfFreqTest()
+      : EncoderTest(GET_PARAM(0)),
+        test_video_param_(GET_PARAM(1)),
+        test_encode_param_(GET_PARAM(2)),
+        min_arf_requested_(GET_PARAM(3)) {
+  }
+
+  virtual ~ArfFreqTest() {}
+
+  virtual void SetUp() {
+    InitializeConfig();
+    SetMode(test_encode_param_.mode);
+    if (test_encode_param_.mode != ::libvpx_test::kRealTime) {
+      cfg_.g_lag_in_frames = 25;
+      cfg_.rc_end_usage = VPX_VBR;
+    } else {
+      cfg_.g_lag_in_frames = 0;
+      cfg_.rc_end_usage = VPX_CBR;
+      cfg_.rc_buf_sz = 1000;
+      cfg_.rc_buf_initial_sz = 500;
+      cfg_.rc_buf_optimal_sz = 600;
+    }
+    dec_cfg_.threads = 4;
+  }
+
+  virtual void BeginPassHook(unsigned int) {
+    min_arf_ = ARF_NOT_SEEN;
+    run_of_visible_frames_ = 0;
+  }
+
+  int GetNumFramesInPkt(const vpx_codec_cx_pkt_t *pkt) {
+    const uint8_t *buffer = reinterpret_cast<uint8_t*>(pkt->data.frame.buf);
+    const uint8_t marker = buffer[pkt->data.frame.sz - 1];
+    const int mag = ((marker >> 3) & 3) + 1;
+    int frames = (marker & 0x7) + 1;
+    const unsigned int index_sz = 2 + mag  * frames;
+    // Check for superframe or not.
+    // Assume superframe has only one visible frame, the rest being
+    // invisible. If superframe index is not found, then there is only
+    // one frame.
+    if (!((marker & 0xe0) == 0xc0 &&
+          pkt->data.frame.sz >= index_sz &&
+          buffer[pkt->data.frame.sz - index_sz] == marker)) {
+      frames = 1;
+    }
+    return frames;
+  }
+
+  virtual void FramePktHook(const vpx_codec_cx_pkt_t *pkt) {
+    if (pkt->kind != VPX_CODEC_CX_FRAME_PKT)
+      return;
+    const int frames = GetNumFramesInPkt(pkt);
+    if (frames == 1) {
+      run_of_visible_frames_++;
+    } else if (frames == 2) {
+      if (min_arf_ == ARF_NOT_SEEN) {
+        min_arf_ = ARF_SEEN_ONCE;
+      } else if (min_arf_ == ARF_SEEN_ONCE ||
+                 run_of_visible_frames_ < min_arf_) {
+        min_arf_ = run_of_visible_frames_;
+      }
+      run_of_visible_frames_ = 1;
+    } else {
+      min_arf_ = 0;
+      run_of_visible_frames_ = 1;
+    }
+  }
+
+  virtual void PreEncodeFrameHook(::libvpx_test::VideoSource *video,
+                                  ::libvpx_test::Encoder *encoder) {
+    if (video->frame() == 0) {
+      encoder->Control(VP9E_SET_FRAME_PARALLEL_DECODING, 1);
+      encoder->Control(VP9E_SET_TILE_COLUMNS, 4);
+      encoder->Control(VP8E_SET_CPUUSED, test_encode_param_.cpu_used);
+      encoder->Control(VP9E_SET_MIN_GF_INTERVAL, min_arf_requested_);
+      if (test_encode_param_.mode != ::libvpx_test::kRealTime) {
+        encoder->Control(VP8E_SET_ENABLEAUTOALTREF, 1);
+        encoder->Control(VP8E_SET_ARNR_MAXFRAMES, 7);
+        encoder->Control(VP8E_SET_ARNR_STRENGTH, 5);
+        encoder->Control(VP8E_SET_ARNR_TYPE, 3);
+      }
+    }
+  }
+
+  int GetMinArfDistance() const {
+    return min_arf_;
+  }
+
+  int GetMinArfDistanceRequested() const {
+    if (min_arf_requested_)
+      return min_arf_requested_;
+    else
+      return vp9_rc_get_default_min_gf_interval(
+          test_video_param_.width, test_video_param_.height,
+          (double)test_video_param_.framerate_num /
+          test_video_param_.framerate_den);
+  }
+
+  TestVideoParam test_video_param_;
+  TestEncodeParam test_encode_param_;
+
+ private:
+  int min_arf_requested_;
+  int min_arf_;
+  int run_of_visible_frames_;
+};
+
+TEST_P(ArfFreqTest, MinArfFreqTest) {
+  cfg_.rc_target_bitrate = kBitrate;
+  cfg_.g_error_resilient = 0;
+  cfg_.g_profile = test_video_param_.profile;
+  cfg_.g_input_bit_depth = test_video_param_.input_bit_depth;
+  cfg_.g_bit_depth = test_video_param_.bit_depth;
+  init_flags_ = VPX_CODEC_USE_PSNR;
+  if (cfg_.g_bit_depth > 8)
+    init_flags_ |= VPX_CODEC_USE_HIGHBITDEPTH;
+
+  libvpx_test::VideoSource *video;
+  if (is_extension_y4m(test_video_param_.filename)) {
+    video = new libvpx_test::Y4mVideoSource(test_video_param_.filename,
+                                            0, kFrames);
+  } else {
+    video = new libvpx_test::YUVVideoSource(test_video_param_.filename,
+                                            test_video_param_.fmt,
+                                            test_video_param_.width,
+                                            test_video_param_.height,
+                                            test_video_param_.framerate_num,
+                                            test_video_param_.framerate_den,
+                                            0, kFrames);
+  }
+
+  ASSERT_NO_FATAL_FAILURE(RunLoop(video));
+  const int min_arf_dist = GetMinArfDistance();
+  const int min_arf_dist_requested = GetMinArfDistanceRequested();
+  if (min_arf_dist != ARF_NOT_SEEN && min_arf_dist != ARF_SEEN_ONCE) {
+    EXPECT_GE(min_arf_dist, min_arf_dist_requested);
+  }
+  delete(video);
+}
+
+VP9_INSTANTIATE_TEST_CASE(
+    ArfFreqTest,
+    ::testing::ValuesIn(kTestVectors),
+    ::testing::ValuesIn(kEncodeVectors),
+    ::testing::ValuesIn(kMinArfVectors));
+}  // namespace

vp9/common/vp9_reconinter.c

@@ -168,9 +168,9 @@ MV average_split_mvs(const struct macroblockd_plane *pd,
 }
 
 void build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
-                                   int bw, int bh,
+                            int bw, int bh,
-                                   int x, int y, int w, int h,
+                            int x, int y, int w, int h,
-                                   int mi_x, int mi_y) {
+                            int mi_x, int mi_y) {
   struct macroblockd_plane *const pd = &xd->plane[plane];
   const MODE_INFO *mi = xd->mi[0];
   const int is_compound = has_second_ref(&mi->mbmi);
@@ -201,7 +201,20 @@ void build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
     const int is_scaled = vp9_is_scaled(sf);
 
     if (is_scaled) {
-      pre = pre_buf->buf + scaled_buffer_offset(x, y, pre_buf->stride, sf);
+      // Co-ordinate of containing block to pixel precision.
+      int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
+      int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
+
+      if (plane == 0)
+        pre_buf->buf = xd->block_refs[ref]->buf->y_buffer;
+      else if (plane == 1)
+        pre_buf->buf = xd->block_refs[ref]->buf->u_buffer;
+      else
+        pre_buf->buf = xd->block_refs[ref]->buf->v_buffer;
+
+      pre_buf->buf += scaled_buffer_offset(x_start + x, y_start + y,
+                                           pre_buf->stride, sf);
+      pre = pre_buf->buf;
       scaled_mv = vp9_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
       xs = sf->x_step_q4;
       ys = sf->y_step_q4;

vp9/decoder/vp9_decodeframe.c

@@ -746,6 +746,12 @@ static void decode_block(VP9Decoder *const pbi, MACROBLOCKD *const xd,
   if (less8x8)
     bsize = BLOCK_8X8;
 
+#if CONFIG_INTERNAL_STATS
+  pbi->total_block_in_8x8 += (1 << (num_pels_log2_lookup[bsize] - 6));
+  if (!is_inter_block(mbmi))
+    pbi->intra_block_in_8x8 += (1 << (num_pels_log2_lookup[bsize] - 6));
+#endif
+
   if (mbmi->skip) {
     reset_skip_context(xd, bsize);
   }
@@ -754,6 +760,9 @@ static void decode_block(VP9Decoder *const pbi, MACROBLOCKD *const xd,
     struct intra_args arg = {xd, r, mbmi->segment_id};
     vp9_foreach_transformed_block(xd, bsize,
                                   predict_and_reconstruct_intra_block, &arg);
+#if CONFIG_INTERNAL_STATS
+    pbi->sub8x8_intra += less8x8;
+#endif
   } else {
     // Prediction
     dec_build_inter_predictors_sb(pbi, xd, mi_row, mi_col, bsize);
@@ -766,6 +775,10 @@ static void decode_block(VP9Decoder *const pbi, MACROBLOCKD *const xd,
       if (!less8x8 && eobtotal == 0)
         mbmi->skip = 1;  // skip loopfilter
     }
+
+#if CONFIG_INTERNAL_STATS
+  pbi->sub8x8_inter += less8x8;
+#endif
   }
 
   xd->corrupted |= vp9_reader_has_error(r);
@@ -1681,6 +1694,7 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
       for (i = 0; i < REF_FRAMES; ++i)
         cm->next_ref_frame_map[i] = cm->ref_frame_map[i];
     }
+
     return 0;
   }
 
@@ -1793,6 +1807,17 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
   // below, forcing the use of context 0 for those frame types.
   cm->frame_context_idx = vp9_rb_read_literal(rb, FRAME_CONTEXTS_LOG2);
 
+#if CONFIG_FULL_BUFFER_TEST
+  {
+    int k;
+    fprintf(stderr, "\nDecode frame index %d, is_show_frame %d, new_fb_idx %d\n",
+            cm->current_video_frame, cm->show_frame, cm->new_fb_idx);
+    for (k = 0; k < 8; ++k)
+      fprintf(stderr, "%d  ", cm->ref_frame_map[k]);
+    fprintf(stderr, "\n");
+  }
+#endif
+
   // Generate next_ref_frame_map.
   lock_buffer_pool(pool);
   for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
@@ -1979,6 +2004,9 @@ void vp9_decode_frame(VP9Decoder *pbi,
       init_read_bit_buffer(pbi, &rb, data, data_end, clear_data));
   const int tile_rows = 1 << cm->log2_tile_rows;
   const int tile_cols = 1 << cm->log2_tile_cols;
+#if CONFIG_INTERNAL_STATS
+  const int64_t data_size = data_end - data;
+#endif
   YV12_BUFFER_CONFIG *const new_fb = get_frame_new_buffer(cm);
   xd->cur_buf = new_fb;
 
@@ -2008,6 +2036,23 @@ void vp9_decode_frame(VP9Decoder *pbi,
                        "Uninitialized entropy context.");
 
   vp9_zero(cm->counts);
+#if CONFIG_INTERNAL_STATS
+  // Reset internal stats
+  if (cm->current_video_frame == 0) {
+    int i;
+    for (i = 0; i < BR_MOVING_AVERAGE_SIZE; ++i)
+      pbi->br_data.bit_rate[i] = 0;
+    pbi->br_data.index = 0;
+    pbi->peak_average_br = 0;
+    pbi->total_block_in_8x8 = 0;
+    pbi->subpel_mc_block_in_4x4_h = 0;
+    pbi->subpel_mc_block_in_4x4_v = 0;
+    pbi->sub8x8_inter = 0;
+    pbi->sub8x8_intra = 0;
+    pbi->intra_block_in_8x8 = 0;
+    pbi->compound_inter_block_in_8x8 = 0;
+  }
+#endif
 
   xd->corrupted = 0;
   new_fb->corrupted = read_compressed_header(pbi, data, first_partition_size);
@@ -2024,6 +2069,7 @@ void vp9_decode_frame(VP9Decoder *pbi,
   if (pbi->frame_parallel_decode && cm->frame_parallel_decoding_mode) {
     VP9Worker *const worker = pbi->frame_worker_owner;
     FrameWorkerData *const frame_worker_data = worker->data1;
+
     if (cm->refresh_frame_context) {
       context_updated = 1;
       cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
@@ -2076,4 +2122,43 @@ void vp9_decode_frame(VP9Decoder *pbi,
   // Non frame parallel update frame context here.
   if (cm->refresh_frame_context && !context_updated)
     cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
+
+#if CONFIG_INTERNAL_STATS
+  vp9_clear_system_state();
+  {
+    FILE *pf = fopen("frame_level_stats.stt", "a");
+    int i;
+    double subpel_mc_h = (double)pbi->subpel_mc_block_in_4x4_h /
+        (double)pbi->total_block_in_8x8;
+    double subpel_mc_v = (double)pbi->subpel_mc_block_in_4x4_v /
+        (double)pbi->total_block_in_8x8;
+    double intra_mode = (double)pbi->intra_block_in_8x8 /
+        (double)pbi->total_block_in_8x8;
+    double compound_block = (double)pbi->compound_inter_block_in_8x8 /
+        (double)pbi->total_block_in_8x8;
+    double sub8x8_inter = (double)pbi->sub8x8_inter /
+        (double)pbi->total_block_in_8x8;
+    double sub8x8_intra = (double)pbi->sub8x8_intra /
+        (double)pbi->total_block_in_8x8;
+    double average_bit_rate = 0;
+
+    pbi->br_data.bit_rate[pbi->br_data.index] = data_size;
+    ++pbi->br_data.index;
+    if (pbi->br_data.index >= BR_MOVING_AVERAGE_SIZE)
+      pbi->br_data.index = 0;
+    for (i = 0; i < BR_MOVING_AVERAGE_SIZE; ++i) {
+      average_bit_rate += (double)pbi->br_data.bit_rate[i];
+    }
+    average_bit_rate /= BR_MOVING_AVERAGE_SIZE;
+    if ((int64_t)average_bit_rate > pbi->peak_average_br)
+      pbi->peak_average_br = (int64_t)average_bit_rate;
+
+    fprintf(pf, "frame index %d, sub-pel mc (%7.3f, %7.3f)\tintra mode%7.3f\tcompound inter block%7.3f\t",
+            cm->current_video_frame, 25 * subpel_mc_h, 25 * subpel_mc_v, 100 * intra_mode,
+            100 * compound_block);
+    fprintf(pf, "sub8x8 inter %7.3f intra %7.3f average-br %7.3f peak-average-br %7.3f\n",
+            100 * sub8x8_inter, 100 * sub8x8_intra, average_bit_rate, (double)pbi->peak_average_br / 1000);
+    fclose(pf);
+  }
+#endif
 }

vp9/decoder/vp9_decodemv.c

@@ -479,6 +479,10 @@ static void read_inter_block_mode_info(VP9Decoder *const pbi,
   read_ref_frames(cm, xd, r, mbmi->segment_id, mbmi->ref_frame);
   is_compound = has_second_ref(mbmi);
 
+#if CONFIG_INTERNAL_STATS
+  pbi->compound_inter_block_in_8x8 += is_compound;
+#endif
+
   for (ref = 0; ref < 1 + is_compound; ++ref) {
     const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
     RefBuffer *ref_buf = &cm->frame_refs[frame - LAST_FRAME];
@@ -547,7 +551,22 @@ static void read_inter_block_mode_info(VP9Decoder *const pbi,
         mi->bmi[j].as_mv[0].as_int = block[0].as_int;
         if (is_compound)
           mi->bmi[j].as_mv[1].as_int = block[1].as_int;
-
+#if CONFIG_INTERNAL_STATS
+        if (mi->bmi[j].as_mv[0].as_mv.row & 0x07)
+          pbi->subpel_mc_block_in_4x4_v +=
+              (1 << (num_pels_log2_lookup[bsize] - 4));
+        if (mi->bmi[j].as_mv[0].as_mv.col & 0x07)
+          pbi->subpel_mc_block_in_4x4_h +=
+              (1 << (num_pels_log2_lookup[bsize] - 4));
+        if (is_compound) {
+          if (mi->bmi[j].as_mv[1].as_mv.row & 0x07)
+            pbi->subpel_mc_block_in_4x4_v +=
+                (1 << (num_pels_log2_lookup[bsize] - 4));
+          if (mi->bmi[j].as_mv[1].as_mv.col & 0x07)
+            pbi->subpel_mc_block_in_4x4_h +=
+                (1 << (num_pels_log2_lookup[bsize] - 4));
+        }
+#endif
         if (num_4x4_h == 2)
           mi->bmi[j + 2] = mi->bmi[j];
         if (num_4x4_w == 2)
@@ -562,6 +581,18 @@ static void read_inter_block_mode_info(VP9Decoder *const pbi,
   } else {
     xd->corrupted |= !assign_mv(cm, xd, mbmi->mode, mbmi->mv, nearestmv,
                                 nearestmv, nearmv, is_compound, allow_hp, r);
+#if CONFIG_INTERNAL_STATS
+    if (mbmi->mv[0].as_mv.row & 0x07)
+      pbi->subpel_mc_block_in_4x4_v += (1 << (num_pels_log2_lookup[bsize] - 4));
+    if (mbmi->mv[0].as_mv.col & 0x07)
+      pbi->subpel_mc_block_in_4x4_h += (1 << (num_pels_log2_lookup[bsize] - 4));
+    if (is_compound) {
+      if (mbmi->mv[1].as_mv.row & 0x07)
+        pbi->subpel_mc_block_in_4x4_v += (1 << (num_pels_log2_lookup[bsize] - 4));
+      if (mbmi->mv[1].as_mv.col & 0x07)
+        pbi->subpel_mc_block_in_4x4_h += (1 << (num_pels_log2_lookup[bsize] - 4));
+    }
+#endif
   }
 }
 

vp9/decoder/vp9_decoder.h

@@ -26,6 +26,14 @@
 extern "C" {
 #endif
 
+#if CONFIG_INTERNAL_STATS
+#define BR_MOVING_AVERAGE_SIZE 4
+typedef struct BR_DATA {
+  int64_t bit_rate[BR_MOVING_AVERAGE_SIZE];
+  int index;
+} BR_DATA;
+#endif
+
 // TODO(hkuang): combine this with TileWorkerData.
 typedef struct TileData {
   VP9_COMMON *cm;
@@ -75,6 +83,31 @@ typedef struct VP9Decoder {
   int inv_tile_order;
   int need_resync;  // wait for key/intra-only frame.
   int hold_ref_buf;  // hold the reference buffer.
+
+#if CONFIG_INTERNAL_STATS
+  // total blocks in unit of 8x8 inside the frame.
+  int64_t total_block_in_8x8;
+
+  // sub8x8 blocks in unit of 8x8.
+  int64_t sub8x8_inter;
+  int64_t sub8x8_intra;
+
+  // number of blocks using 1-D sub-pixel filtering for motion compensated
+  // prediction. if a block is using sub-pixel filter in both vertical and
+  // horizontal directions, it counts as using 1-D sub-pixel filter twice.
+  int64_t subpel_mc_block_in_4x4_h;
+  int64_t subpel_mc_block_in_4x4_v;
+
+  // nubmer of blocks using intra prediction mode.
+  int64_t intra_block_in_8x8;
+
+  // number of blocks using compound prediction mode.
+  int64_t compound_inter_block_in_8x8;
+
+  // peak N-frame-average bit rate
+  int64_t peak_average_br;
+  BR_DATA br_data;
+#endif
 } VP9Decoder;
 
 int vp9_receive_compressed_data(struct VP9Decoder *pbi,

vp9/encoder/vp9_bitstream.c

@@ -897,6 +897,21 @@ static void write_tile_info(const VP9_COMMON *const cm,
 }
 
 static int get_refresh_mask(VP9_COMP *cpi) {
+  int lst_fb_idx = cpi->lst_fb_idx;
+
+#if CONFIG_FULL_BUFFER_TEST
+  if (cpi->common.frame_type == KEY_FRAME) {
+    lst_fb_idx = 0;
+  } else {
+    if (cpi->lst_fb_idx == 0)
+      lst_fb_idx = 3;
+    else if (cpi->lst_fb_idx == 7)
+      lst_fb_idx = 0;
+    else
+      lst_fb_idx = cpi->lst_fb_idx + 1;
+  }
+#endif
+
   if (vp9_preserve_existing_gf(cpi)) {
     // We have decided to preserve the previously existing golden frame as our
     // new ARF frame. However, in the short term we leave it in the GF slot and,
@@ -908,7 +923,7 @@ static int get_refresh_mask(VP9_COMP *cpi) {
     // Note: This is highly specific to the use of ARF as a forward reference,
     // and this needs to be generalized as other uses are implemented
     // (like RTC/temporal scalability).
-    return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
+    return (cpi->refresh_last_frame << lst_fb_idx) |
            (cpi->refresh_golden_frame << cpi->alt_fb_idx);
   } else {
     int arf_idx = cpi->alt_fb_idx;
@@ -916,7 +931,7 @@ static int get_refresh_mask(VP9_COMP *cpi) {
       const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
       arf_idx = gf_group->arf_update_idx[gf_group->index];
     }
-    return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
+    return (cpi->refresh_last_frame << lst_fb_idx) |
            (cpi->refresh_golden_frame << cpi->gld_fb_idx) |
            (cpi->refresh_alt_ref_frame << arf_idx);
   }

vp9/encoder/vp9_encoder.c

@@ -1447,7 +1447,7 @@ void vp9_change_config(struct VP9_COMP *cpi, const VP9EncoderConfig *oxcf) {
   cpi->td.mb.e_mbd.bd = (int)cm->bit_depth;
 #endif  // CONFIG_VP9_HIGHBITDEPTH
 
-  rc->baseline_gf_interval = DEFAULT_GF_INTERVAL;
+  rc->baseline_gf_interval = (MIN_GF_INTERVAL + MAX_GF_INTERVAL) / 2;
 
   cpi->refresh_golden_frame = 0;
   cpi->refresh_last_frame = 1;
@@ -2697,13 +2697,61 @@ void vp9_update_reference_frames(VP9_COMP *cpi) {
   }
 
   if (cpi->refresh_last_frame) {
+#if CONFIG_FULL_BUFFER_TEST
+    int ref_index;
+
+    if (cm->frame_type == KEY_FRAME) {
+      cpi->lst_fb_idx = 0;
+      ref_index = cm->ref_frame_map[cpi->lst_fb_idx];
+      if (ref_index >= 0 && pool->frame_bufs[ref_index].ref_count > 0)
+        pool->frame_bufs[ref_index].ref_count--;
+
+      cm->ref_frame_map[cpi->lst_fb_idx] = cm->new_fb_idx;
+
+      pool->frame_bufs[cm->new_fb_idx].ref_count++;
+    } else {
+      if (cpi->lst_fb_idx == 0)
+        cpi->lst_fb_idx = 3;
+      else if (cpi->lst_fb_idx == 7)
+        cpi->lst_fb_idx = 0;
+      else
+        ++cpi->lst_fb_idx;
+
+      ref_index = cm->ref_frame_map[cpi->lst_fb_idx];
+      if (ref_index >= 0 && pool->frame_bufs[ref_index].ref_count > 0)
+        pool->frame_bufs[ref_index].ref_count--;
+
+      ref_index = cm->ref_frame_map[cpi->lst_fb_idx];
+      cm->ref_frame_map[cpi->lst_fb_idx] = cm->new_fb_idx;
+      pool->frame_bufs[cm->new_fb_idx].ref_count++;
+    }
+#else
     ref_cnt_fb(pool->frame_bufs,
                &cm->ref_frame_map[cpi->lst_fb_idx], cm->new_fb_idx);
+#endif
+
     if (!cpi->rc.is_src_frame_alt_ref)
       memcpy(cpi->interp_filter_selected[LAST_FRAME],
              cpi->interp_filter_selected[0],
              sizeof(cpi->interp_filter_selected[0]));
   }
+
+#if CONFIG_FULL_BUFFER_TEST
+//  {
+//    int k;
+//    fprintf(stderr, "\nEecode frame index %d, is_show_frame %d, new_fb_idx %d\n",
+//            cm->current_video_frame, cm->show_frame, cm->new_fb_idx);
+//    for (k = 0; k < 8; ++k)
+//      fprintf(stderr, "%d  ", cm->ref_frame_map[k]);
+//    fprintf(stderr, "\n");
+//    if (cm->frame_type == KEY_FRAME)
+//      fprintf(stderr, "key frame\n\n");
+//
+//    for (k = 0; k < 16; ++k)
+//      fprintf(stderr, "%d  ", pool->frame_bufs[cm->new_fb_idx].buf.y_buffer[k]);
+//  }
+#endif
+
 #if CONFIG_VP9_TEMPORAL_DENOISING
   if (cpi->oxcf.noise_sensitivity > 0) {
     vp9_denoiser_update_frame_info(&cpi->denoiser,
@@ -2786,7 +2834,6 @@ void vp9_scale_references(VP9_COMP *cpi) {
         if (cm->new_fb_idx == INVALID_IDX)
           return;
         new_fb_ptr = &pool->frame_bufs[new_fb];
-        cm->cur_frame = &pool->frame_bufs[new_fb];
         vp9_realloc_frame_buffer(&pool->frame_bufs[new_fb].buf,
                                  cm->width, cm->height,
                                  cm->subsampling_x, cm->subsampling_y,
@@ -3023,6 +3070,17 @@ static void set_frame_size(VP9_COMP *cpi) {
   VP9EncoderConfig *const oxcf = &cpi->oxcf;
   MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
 
+#if CONFIG_INTERNAL_RESIZE
+  if (oxcf->pass == 2 && oxcf->resize_mode == RESIZE_DYNAMIC) {
+    if ((cm->current_video_frame % 15) == 3) {
+      cpi->resize_pending = 1;
+      cpi->rc.frame_size_selector = (cm->current_video_frame / 15) & 0x01;
+    } else {
+      cpi->resize_pending = 0;
+    }
+  }
+#endif
+
   if (oxcf->pass == 2 &&
       oxcf->rc_mode == VPX_VBR &&
       ((oxcf->resize_mode == RESIZE_FIXED && cm->current_video_frame == 0) ||
@@ -3035,6 +3093,11 @@ static void set_frame_size(VP9_COMP *cpi) {
                          oxcf->scaled_frame_height);
   }
 
+#if CONFIG_INTERNAL_RESIZE
+  fprintf(stderr, "%d %d\n", oxcf->scaled_frame_width,
+          oxcf->scaled_frame_height);
+#endif
+
   if (oxcf->pass == 0 &&
       oxcf->rc_mode == VPX_CBR &&
       !cpi->use_svc &&
@@ -4197,6 +4260,26 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
     set_frame_size(cpi);
   }
 
+#if CONFIG_FULL_BUFFER_TEST
+  if (cm->current_video_frame > 12 && oxcf->pass == 2) {
+    int mod_idx = cm->current_video_frame & 0x3F;
+    if (mod_idx > 12 && mod_idx < 20) {
+      cpi->refresh_last_frame = 0;
+
+      if (cpi->common.frame_type == KEY_FRAME) {
+        cpi->lst_fb_idx = 0;
+      } else {
+        if (cpi->lst_fb_idx == 0)
+          cpi->lst_fb_idx = 3;
+        else if (cpi->lst_fb_idx == 7)
+          cpi->lst_fb_idx = 0;
+        else
+          ++cpi->lst_fb_idx;
+      }
+    }
+  }
+#endif
+
   for (i = 0; i < MAX_REF_FRAMES; ++i)
     cpi->scaled_ref_idx[i] = INVALID_IDX;
 

vp9/encoder/vp9_encoder.h

@@ -50,8 +50,6 @@
 extern "C" {
 #endif
 
-#define DEFAULT_GF_INTERVAL         10
-
 typedef struct {
   int nmvjointcost[MV_JOINTS];
   int nmvcosts[2][MV_VALS];
@@ -219,6 +217,9 @@ typedef struct VP9EncoderConfig {
   int arnr_max_frames;
   int arnr_strength;
 
+  int min_gf_interval;
+  int max_gf_interval;
+
   int tile_columns;
   int tile_rows;
 

vp9/encoder/vp9_firstpass.c

@@ -1849,7 +1849,8 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
   int64_t gf_group_bits;
   double gf_group_error_left;
   int gf_arf_bits;
-  int is_key_frame = frame_is_intra_only(cm);
+  const int is_key_frame = frame_is_intra_only(cm);
+  const int kf_or_arf_active = is_key_frame || rc->source_alt_ref_active;
 
   // Reset the GF group data structures unless this is a key
   // frame in which case it will already have been done.
@@ -1902,7 +1903,10 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
       // bits to spare and are better with a smaller interval and smaller boost.
       // At high Q when there are few bits to spare we are better with a longer
       // interval to spread the cost of the GF.
-      active_max_gf_interval = 12 + MIN(4, (int_lbq / 6));
+      active_max_gf_interval = rc->max_gf_interval - 4 + MIN(4, (int_lbq / 6));
+      if (active_max_gf_interval < active_min_gf_interval)
+        active_max_gf_interval = active_min_gf_interval;
+
       if (active_max_gf_interval > rc->max_gf_interval)
         active_max_gf_interval = rc->max_gf_interval;
       if (active_max_gf_interval < active_min_gf_interval)
@@ -1964,10 +1968,11 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
     // Break out conditions.
     if (
       // Break at active_max_gf_interval unless almost totally static.
-      (i >= active_max_gf_interval && (zero_motion_accumulator < 0.995)) ||
+      ((i >= active_max_gf_interval + kf_or_arf_active) &&
+       (zero_motion_accumulator < 0.995)) ||
       (
         // Don't break out with a very short interval.
-        (i > active_min_gf_interval) &&
+        (i >= active_min_gf_interval + kf_or_arf_active) &&
         (!flash_detected) &&
         ((mv_ratio_accumulator > mv_ratio_accumulator_thresh) ||
          (abs_mv_in_out_accumulator > 3.0) ||

vp9/encoder/vp9_ratectrl.c

@@ -276,6 +276,27 @@ static void update_buffer_level(VP9_COMP *cpi, int encoded_frame_size) {
   }
 }
 
+int vp9_rc_get_default_min_gf_interval(
+    int width, int height, double framerate) {
+  // Assume we do not need any constraint lower than 4K 20 fps
+  static const double factor_safe = 3840 * 2160 * 20.0;
+  const double factor = width * height * framerate;
+
+  if (factor <= factor_safe)
+    return MIN_GF_INTERVAL;
+  else
+    return (int)(MIN_GF_INTERVAL * factor / factor_safe + 0.5);
+  // Note this logic makes:
+  // 4K24: 5
+  // 4K30: 6
+  // 4K60: 12
+}
+
+int vp9_rc_get_default_max_gf_interval(double framerate, int min_gf_interval) {
+  int interval = MIN(MAX_GF_INTERVAL, (int)(framerate * 0.75));
+  return MAX(interval, min_gf_interval);
+}
+
 void vp9_rc_init(const VP9EncoderConfig *oxcf, int pass, RATE_CONTROL *rc) {
   int i;
 
@@ -284,9 +305,9 @@ void vp9_rc_init(const VP9EncoderConfig *oxcf, int pass, RATE_CONTROL *rc) {
     rc->avg_frame_qindex[INTER_FRAME] = oxcf->worst_allowed_q;
   } else {
     rc->avg_frame_qindex[KEY_FRAME] = (oxcf->worst_allowed_q +
-                                           oxcf->best_allowed_q) / 2;
+                                       oxcf->best_allowed_q) / 2;
     rc->avg_frame_qindex[INTER_FRAME] = (oxcf->worst_allowed_q +
-                                           oxcf->best_allowed_q) / 2;
+                                         oxcf->best_allowed_q) / 2;
   }
 
   rc->last_q[KEY_FRAME] = oxcf->best_allowed_q;
@@ -304,7 +325,6 @@ void vp9_rc_init(const VP9EncoderConfig *oxcf, int pass, RATE_CONTROL *rc) {
   rc->total_target_bits = 0;
   rc->total_target_vs_actual = 0;
 
-  rc->baseline_gf_interval = DEFAULT_GF_INTERVAL;
   rc->frames_since_key = 8;  // Sensible default for first frame.
   rc->this_key_frame_forced = 0;
   rc->next_key_frame_forced = 0;
@@ -322,6 +342,16 @@ void vp9_rc_init(const VP9EncoderConfig *oxcf, int pass, RATE_CONTROL *rc) {
   for (i = 0; i < RATE_FACTOR_LEVELS; ++i) {
     rc->rate_correction_factors[i] = 1.0;
   }
+
+  rc->min_gf_interval = oxcf->min_gf_interval;
+  rc->max_gf_interval = oxcf->max_gf_interval;
+  if (rc->min_gf_interval == 0)
+    rc->min_gf_interval = vp9_rc_get_default_min_gf_interval(
+        oxcf->width, oxcf->height, oxcf->init_framerate);
+  if (rc->max_gf_interval == 0)
+    rc->max_gf_interval = vp9_rc_get_default_max_gf_interval(
+        oxcf->init_framerate, rc->min_gf_interval);
+  rc->baseline_gf_interval = (rc->min_gf_interval + rc->max_gf_interval) / 2;
 }
 
 int vp9_rc_drop_frame(VP9_COMP *cpi) {
@@ -1382,7 +1412,7 @@ void vp9_rc_get_one_pass_vbr_params(VP9_COMP *cpi) {
     cm->frame_type = INTER_FRAME;
   }
   if (rc->frames_till_gf_update_due == 0) {
-    rc->baseline_gf_interval = DEFAULT_GF_INTERVAL;
+    rc->baseline_gf_interval = (rc->min_gf_interval + rc->max_gf_interval) / 2;
     rc->frames_till_gf_update_due = rc->baseline_gf_interval;
     // NOTE: frames_till_gf_update_due must be <= frames_to_key.
     if (rc->frames_till_gf_update_due > rc->frames_to_key) {
@@ -1576,7 +1606,8 @@ void vp9_rc_get_one_pass_cbr_params(VP9_COMP *cpi) {
     if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
       vp9_cyclic_refresh_set_golden_update(cpi);
     else
-      rc->baseline_gf_interval = DEFAULT_GF_INTERVAL;
+      rc->baseline_gf_interval =
+          (rc->min_gf_interval + rc->max_gf_interval) / 2;
     rc->frames_till_gf_update_due = rc->baseline_gf_interval;
     // NOTE: frames_till_gf_update_due must be <= frames_to_key.
     if (rc->frames_till_gf_update_due > rc->frames_to_key)
@@ -1649,20 +1680,19 @@ int vp9_compute_qdelta_by_rate(const RATE_CONTROL *rc, FRAME_TYPE frame_type,
   return target_index - qindex;
 }
 
-#define MIN_GF_INTERVAL     4
-#define MAX_GF_INTERVAL     16
 void vp9_rc_set_gf_interval_range(const VP9_COMP *const cpi,
                                   RATE_CONTROL *const rc) {
   const VP9EncoderConfig *const oxcf = &cpi->oxcf;
 
-  // Set a minimum interval.
+  // Set Maximum gf/arf interval
-  rc->min_gf_interval =
+  rc->max_gf_interval = oxcf->max_gf_interval;
-    MIN(MAX_GF_INTERVAL, MAX(MIN_GF_INTERVAL, (int)(cpi->framerate * 0.125)));
+  rc->min_gf_interval = oxcf->min_gf_interval;
-
+  if (rc->min_gf_interval == 0)
-  // Set Maximum gf/arf interval.
+    rc->min_gf_interval = vp9_rc_get_default_min_gf_interval(
-  rc->max_gf_interval =
+        oxcf->width, oxcf->height, cpi->framerate);
-    MIN(MAX_GF_INTERVAL, (int)(cpi->framerate * 0.75));
+  if (rc->max_gf_interval == 0)
-  // Round up to next even number if odd.
+    rc->max_gf_interval = vp9_rc_get_default_max_gf_interval(
+        cpi->framerate, rc->min_gf_interval);
   rc->max_gf_interval += (rc->max_gf_interval & 0x01);
 
   // Extended interval for genuinely static scenes

vp9/encoder/vp9_ratectrl.h

@@ -24,6 +24,9 @@ extern "C" {
 // Bits Per MB at different Q (Multiplied by 512)
 #define BPER_MB_NORMBITS    9
 
+#define MIN_GF_INTERVAL     4
+#define MAX_GF_INTERVAL     16
+
 typedef enum {
   INTER_NORMAL = 0,
   INTER_HIGH = 1,
@@ -155,6 +158,12 @@ double vp9_convert_qindex_to_q(int qindex, vpx_bit_depth_t bit_depth);
 
 void vp9_rc_init_minq_luts(void);
 
+int vp9_rc_get_default_min_gf_interval(int width, int height, double framerate);
+// Note vp9_rc_get_default_max_gf_interval() requires the min_gf_interval to
+// be passed in to ensure that the max_gf_interval returned is at least as bis
+// as that.
+int vp9_rc_get_default_max_gf_interval(double framerate, int min_frame_rate);
+
 // Generally at the high level, the following flow is expected
 // to be enforced for rate control:
 // First call per frame, one of:

vp9/encoder/vp9_rdopt.c

@@ -1359,11 +1359,25 @@ static int64_t encode_inter_mb_segment(VP9_COMP *cpi,
   const InterpKernel *kernel = vp9_get_interp_kernel(mi->mbmi.interp_filter);
 
   for (ref = 0; ref < 1 + is_compound; ++ref) {
-    const uint8_t *pre = &pd->pre[ref].buf[vp9_raster_block_offset(BLOCK_8X8, i,
+    const int bw = b_width_log2_lookup[BLOCK_8X8];
-                                               pd->pre[ref].stride)];
+    const int h = 4 * (i >> bw);
+    const int w = 4 * (i & ((1 << bw) - 1));
+    const struct scale_factors *sf = &xd->block_refs[ref]->sf;
+    int y_stride = pd->pre[ref].stride;
+    uint8_t *pre = pd->pre[ref].buf + (h * pd->pre[ref].stride + w);
+
+    if (vp9_is_scaled(sf)) {
+      const int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
+      const int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
+
+      y_stride = xd->block_refs[ref]->buf->y_stride;
+      pre = xd->block_refs[ref]->buf->y_buffer;
+      pre += scaled_buffer_offset(x_start + w, y_start + h,
+                                  y_stride, sf);
+    }
 #if CONFIG_VP9_HIGHBITDEPTH
   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
-    vp9_highbd_build_inter_predictor(pre, pd->pre[ref].stride,
+    vp9_highbd_build_inter_predictor(pre, y_stride,
                                      dst, pd->dst.stride,
                                      &mi->bmi[i].as_mv[ref].as_mv,
                                      &xd->block_refs[ref]->sf, width, height,
@@ -1371,7 +1385,7 @@ static int64_t encode_inter_mb_segment(VP9_COMP *cpi,
                                      mi_col * MI_SIZE + 4 * (i % 2),
                                      mi_row * MI_SIZE + 4 * (i / 2), xd->bd);
   } else {
-    vp9_build_inter_predictor(pre, pd->pre[ref].stride,
+    vp9_build_inter_predictor(pre, y_stride,
                               dst, pd->dst.stride,
                               &mi->bmi[i].as_mv[ref].as_mv,
                               &xd->block_refs[ref]->sf, width, height, ref,
@@ -1380,7 +1394,7 @@ static int64_t encode_inter_mb_segment(VP9_COMP *cpi,
                               mi_row * MI_SIZE + 4 * (i / 2));
   }
 #else
-    vp9_build_inter_predictor(pre, pd->pre[ref].stride,
+    vp9_build_inter_predictor(pre, y_stride,
                               dst, pd->dst.stride,
                               &mi->bmi[i].as_mv[ref].as_mv,
                               &xd->block_refs[ref]->sf, width, height, ref,
@@ -3803,6 +3817,13 @@ void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi,
     int64_t total_sse = INT_MAX;
     int early_term = 0;
 
+    struct buf_2d backup_yv12[2][MAX_MB_PLANE];
+    const YV12_BUFFER_CONFIG *scaled_ref_frame[2] = {
+        vp9_get_scaled_ref_frame(cpi, vp9_ref_order[ref_index].ref_frame[0]),
+        vp9_get_scaled_ref_frame(cpi, vp9_ref_order[ref_index].ref_frame[1])
+    };
+    int ref;
+
     ref_frame = vp9_ref_order[ref_index].ref_frame[0];
     second_ref_frame = vp9_ref_order[ref_index].ref_frame[1];
 
@@ -3858,16 +3879,6 @@ void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi,
         continue;
     }
 
-    // TODO(jingning, jkoleszar): scaling reference frame not supported for
-    // sub8x8 blocks.
-    if (ref_frame > INTRA_FRAME &&
-        vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
-      continue;
-
-    if (second_ref_frame > INTRA_FRAME &&
-        vp9_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf))
-      continue;
-
     if (comp_pred)
       mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
     else if (ref_frame != INTRA_FRAME)
@@ -3946,6 +3957,19 @@ void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi,
       int pred_exists = 0;
       int uv_skippable;
 
+      for (ref = 0; ref < 2; ++ref) {
+        if (scaled_ref_frame[ref]) {
+          int i;
+          // Swap out the reference frame for a version that's been scaled to
+          // match the resolution of the current frame, allowing the existing
+          // motion search code to be used without additional modifications.
+          for (i = 0; i < MAX_MB_PLANE; i++)
+            backup_yv12[ref][i] = xd->plane[i].pre[ref];
+          vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
+                               NULL);
+        }
+      }
+
       this_rd_thresh = (ref_frame == LAST_FRAME) ?
           rd_opt->threshes[segment_id][bsize][THR_LAST] :
           rd_opt->threshes[segment_id][bsize][THR_ALTR];
@@ -4079,8 +4103,17 @@ void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi,
                                         BLOCK_8X8);
         memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
         if (!super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable,
-                              &uv_sse, BLOCK_8X8, tmp_best_rdu))
+                              &uv_sse, BLOCK_8X8, tmp_best_rdu)) {
+          for (ref = 0; ref < 2; ++ref) {
+            if (scaled_ref_frame[ref]) {
+              // Restore the prediction frame pointers to their unscaled versions.
+              int i;
+              for (i = 0; i < MAX_MB_PLANE; ++i)
+                xd->plane[i].pre[ref] = backup_yv12[ref][i];
+            }
+          }
           continue;
+        }
 
         rate2 += rate_uv;
         distortion2 += distortion_uv;
@@ -4089,6 +4122,15 @@ void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi,
       }
     }
 
+    for (ref = 0; ref < 2; ++ref) {
+      if (scaled_ref_frame[ref]) {
+        // Restore the prediction frame pointers to their unscaled versions.
+        int i;
+        for (i = 0; i < MAX_MB_PLANE; ++i)
+          xd->plane[i].pre[ref] = backup_yv12[ref][i];
+      }
+    }
+
     if (cm->reference_mode == REFERENCE_MODE_SELECT)
       rate2 += compmode_cost;
 

vp9/encoder/vp9_speed_features.c

@@ -48,8 +48,9 @@ static void set_good_speed_feature_framesize_dependent(VP9_COMP *cpi,
 
   if (speed >= 1) {
     if (MIN(cm->width, cm->height) >= 720) {
-      sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
+//      sf->disable_split_mask = cm->show_frame ? DISABLE_ALL_SPLIT
-                                              : DISABLE_ALL_INTER_SPLIT;
+//                                              : DISABLE_ALL_INTER_SPLIT;
+//      sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
       sf->partition_search_breakout_dist_thr = (1 << 23);
     } else {
       sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;

vp9/vp9_cx_iface.c

@@ -31,6 +31,8 @@ struct vp9_extracfg {
   unsigned int                tile_rows;
   unsigned int                arnr_max_frames;
   unsigned int                arnr_strength;
+  unsigned int                min_gf_interval;
+  unsigned int                max_gf_interval;
   vp8e_tuning                 tuning;
   unsigned int                cq_level;  // constrained quality level
   unsigned int                rc_max_intra_bitrate_pct;
@@ -55,6 +57,8 @@ static struct vp9_extracfg default_extra_cfg = {
   0,                          // tile_rows
   7,                          // arnr_max_frames
   5,                          // arnr_strength
+  0,                          // min_gf_interval; 0 -> default decision
+  0,                          // max_gf_interval; 0 -> default decision
   VP8_TUNE_PSNR,              // tuning
   10,                         // cq_level
   0,                          // rc_max_intra_bitrate_pct
@@ -167,6 +171,12 @@ static vpx_codec_err_t validate_config(vpx_codec_alg_priv_t *ctx,
   RANGE_CHECK_HI(cfg, rc_resize_up_thresh,   100);
   RANGE_CHECK_HI(cfg, rc_resize_down_thresh, 100);
   RANGE_CHECK(cfg,        g_pass,         VPX_RC_ONE_PASS, VPX_RC_LAST_PASS);
+  RANGE_CHECK(extra_cfg, min_gf_interval, 0, (MAX_LAG_BUFFERS - 1));
+  RANGE_CHECK(extra_cfg, max_gf_interval, 0, (MAX_LAG_BUFFERS - 1));
+  if (extra_cfg->min_gf_interval > 0 && extra_cfg->max_gf_interval > 0) {
+    RANGE_CHECK(extra_cfg, max_gf_interval, extra_cfg->min_gf_interval,
+                (MAX_LAG_BUFFERS - 1));
+  }
 
   if (cfg->rc_resize_allowed == 1) {
     RANGE_CHECK(cfg, rc_scaled_width, 0, cfg->g_w);
@@ -454,6 +464,8 @@ static vpx_codec_err_t set_encoder_config(
   oxcf->color_space = extra_cfg->color_space;
   oxcf->arnr_max_frames = extra_cfg->arnr_max_frames;
   oxcf->arnr_strength   = extra_cfg->arnr_strength;
+  oxcf->min_gf_interval = extra_cfg->min_gf_interval;
+  oxcf->max_gf_interval = extra_cfg->max_gf_interval;
 
   oxcf->tuning = extra_cfg->tuning;
   oxcf->content = extra_cfg->content;
@@ -727,6 +739,20 @@ static vpx_codec_err_t ctrl_set_aq_mode(vpx_codec_alg_priv_t *ctx,
   return update_extra_cfg(ctx, &extra_cfg);
 }
 
+static vpx_codec_err_t ctrl_set_min_gf_interval(vpx_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.min_gf_interval = CAST(VP9E_SET_MIN_GF_INTERVAL, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
+static vpx_codec_err_t ctrl_set_max_gf_interval(vpx_codec_alg_priv_t *ctx,
+                                                va_list args) {
+  struct vp9_extracfg extra_cfg = ctx->extra_cfg;
+  extra_cfg.max_gf_interval = CAST(VP9E_SET_MAX_GF_INTERVAL, args);
+  return update_extra_cfg(ctx, &extra_cfg);
+}
+
 static vpx_codec_err_t ctrl_set_frame_periodic_boost(vpx_codec_alg_priv_t *ctx,
                                                      va_list args) {
   struct vp9_extracfg extra_cfg = ctx->extra_cfg;
@@ -1444,6 +1470,8 @@ static vpx_codec_ctrl_fn_map_t encoder_ctrl_maps[] = {
   {VP9E_SET_TUNE_CONTENT,             ctrl_set_tune_content},
   {VP9E_SET_COLOR_SPACE,              ctrl_set_color_space},
   {VP9E_SET_NOISE_SENSITIVITY,        ctrl_set_noise_sensitivity},
+  {VP9E_SET_MIN_GF_INTERVAL,          ctrl_set_min_gf_interval},
+  {VP9E_SET_MAX_GF_INTERVAL,          ctrl_set_max_gf_interval},
 
   // Getters
   {VP8E_GET_LAST_QUANTIZER,           ctrl_get_quantizer},

vpx/vp8cx.h

@@ -518,6 +518,22 @@ enum vp8e_enc_control_id {
    */
   VP9E_SET_TEMPORAL_LAYERING_MODE,
 
+  /*!\brief Codec control function to set minimum interval between GF/ARF frames
+   *
+   * By default the value is set as 4.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_MIN_GF_INTERVAL,
+
+  /*!\brief Codec control function to set minimum interval between GF/ARF frames
+   *
+   * By default the value is set as 16.
+   *
+   * Supported in codecs: VP9
+   */
+  VP9E_SET_MAX_GF_INTERVAL,
+
   /*!\brief Codec control function to get an Active map back from the encoder.
    *
    * Supported in codecs: VP9
@@ -716,6 +732,10 @@ VPX_CTRL_USE_TYPE(VP9E_SET_TUNE_CONTENT, int) /* vp9e_tune_content */
 
 VPX_CTRL_USE_TYPE(VP9E_SET_COLOR_SPACE, int)
 
+VPX_CTRL_USE_TYPE(VP9E_SET_MIN_GF_INTERVAL,  unsigned int)
+
+VPX_CTRL_USE_TYPE(VP9E_SET_MAX_GF_INTERVAL,  unsigned int)
+
 VPX_CTRL_USE_TYPE(VP9E_GET_ACTIVEMAP, vpx_active_map_t *)
 /*! @} - end defgroup vp8_encoder */
 #ifdef __cplusplus

vpxenc.c

@@ -396,6 +396,12 @@ static const arg_def_t gf_cbr_boost_pct = ARG_DEF(
     NULL, "gf-cbr-boost", 1, "Boost for Golden Frame in CBR mode (pct)");
 static const arg_def_t max_inter_rate_pct = ARG_DEF(
     NULL, "max-inter-rate", 1, "Max P-frame bitrate (pct)");
+static const arg_def_t min_gf_interval = ARG_DEF(
+    NULL, "min-gf-interval", 1,
+    "min gf/arf frame interval (default 0, indicating in-built behavior)");
+static const arg_def_t max_gf_interval = ARG_DEF(
+    NULL, "max-gf-interval", 1,
+    "max gf/arf frame interval (default 0, indicating in-built behavior)");
 
 static const struct arg_enum_list color_space_enum[] = {
   { "unknown", VPX_CS_UNKNOWN },
@@ -445,6 +451,7 @@ static const arg_def_t *vp9_args[] = {
   &gf_cbr_boost_pct, &lossless,
   &frame_parallel_decoding, &aq_mode, &frame_periodic_boost,
   &noise_sens, &tune_content, &input_color_space,
+  &min_gf_interval, &max_gf_interval,
 #if CONFIG_VP9 && CONFIG_VP9_HIGHBITDEPTH
   &bitdeptharg, &inbitdeptharg,
 #endif
@@ -460,6 +467,7 @@ static const int vp9_arg_ctrl_map[] = {
   VP9E_SET_LOSSLESS, VP9E_SET_FRAME_PARALLEL_DECODING, VP9E_SET_AQ_MODE,
   VP9E_SET_FRAME_PERIODIC_BOOST, VP9E_SET_NOISE_SENSITIVITY,
   VP9E_SET_TUNE_CONTENT, VP9E_SET_COLOR_SPACE,
+  VP9E_SET_MIN_GF_INTERVAL, VP9E_SET_MAX_GF_INTERVAL,
   0
 };
 #endif