VP9: Add speed 9 for subpel search.

Set subpel search stop to 2 when motion vector is non zero.

10% speedup on 1 and 2 threads on Samsung Galaxy S8+.

Change-Id: I7323bb913000229cf60a37495bf88bcc51d0ac96

9 files changed, 48 insertions, 14 deletions

diff --git a/test/cpu_speed_test.cc b/test/cpu_speed_test.cc
index 404b5b44f..34e35b065 100644
--- a/test/cpu_speed_test.cc
+++ b/test/cpu_speed_test.cc
@@ -152,5 +152,5 @@ VP9_INSTANTIATE_TEST_CASE(CpuSpeedTest,
                           ::testing::Values(::libvpx_test::kTwoPassGood,
                                             ::libvpx_test::kOnePassGood,
                                             ::libvpx_test::kRealTime),
-                          ::testing::Range(0, 9));
+                          ::testing::Range(0, 10));
 }  // namespace
diff --git a/test/encode_perf_test.cc b/test/encode_perf_test.cc
index 0bb435502..142d9e2da 100644
--- a/test/encode_perf_test.cc
+++ b/test/encode_perf_test.cc
@@ -48,7 +48,7 @@ const EncodePerfTestVideo kVP9EncodePerfTestVectors[] = {
   EncodePerfTestVideo("niklas_1280_720_30.yuv", 1280, 720, 600, 470),
 };
 
-const int kEncodePerfTestSpeeds[] = { 5, 6, 7, 8 };
+const int kEncodePerfTestSpeeds[] = { 5, 6, 7, 8, 9 };
 const int kEncodePerfTestThreads[] = { 1, 2, 4 };
 
 #define NELEMENTS(x) (sizeof((x)) / sizeof((x)[0]))
diff --git a/test/svc_datarate_test.cc b/test/svc_datarate_test.cc
index 608f27ebf..5156db750 100644
--- a/test/svc_datarate_test.cc
+++ b/test/svc_datarate_test.cc
@@ -1150,20 +1150,21 @@ TEST_P(DatarateOnePassCbrSvcSmallKF, OnePassCbrSvc2SL3TLSmallKf) {
 }
 
 VP9_INSTANTIATE_TEST_CASE(DatarateOnePassCbrSvcSingleBR,
-                          ::testing::Range(5, 9));
+                          ::testing::Range(5, 10));
 
-VP9_INSTANTIATE_TEST_CASE(DatarateOnePassCbrSvcMultiBR, ::testing::Range(5, 9),
+VP9_INSTANTIATE_TEST_CASE(DatarateOnePassCbrSvcMultiBR, ::testing::Range(5, 10),
                           ::testing::Range(0, 3));
 
 VP9_INSTANTIATE_TEST_CASE(DatarateOnePassCbrSvcFrameDropMultiBR,
-                          ::testing::Range(5, 9), ::testing::Range(0, 2),
+                          ::testing::Range(5, 10), ::testing::Range(0, 2),
                           ::testing::Range(0, 3));
 
 #if CONFIG_VP9_TEMPORAL_DENOISING
-VP9_INSTANTIATE_TEST_CASE(DatarateOnePassCbrSvcDenoiser, ::testing::Range(5, 9),
-                          ::testing::Range(1, 3), ::testing::Range(0, 3));
+VP9_INSTANTIATE_TEST_CASE(DatarateOnePassCbrSvcDenoiser,
+                          ::testing::Range(5, 10), ::testing::Range(1, 3),
+                          ::testing::Range(0, 3));
 #endif
 
-VP9_INSTANTIATE_TEST_CASE(DatarateOnePassCbrSvcSmallKF, ::testing::Range(5, 9),
+VP9_INSTANTIATE_TEST_CASE(DatarateOnePassCbrSvcSmallKF, ::testing::Range(5, 10),
                           ::testing::Range(32, 36));
 }  // namespace
diff --git a/test/vp9_datarate_test.cc b/test/vp9_datarate_test.cc
index c4dbcacbe..cb3aad839 100644
--- a/test/vp9_datarate_test.cc
+++ b/test/vp9_datarate_test.cc
@@ -824,16 +824,17 @@ TEST_P(DatarateTestVP9LargeDenoiser, DenoiserOffOn) {
 VP9_INSTANTIATE_TEST_CASE(DatarateTestVP9Large,
                           ::testing::Values(::libvpx_test::kOnePassGood,
                                             ::libvpx_test::kRealTime),
-                          ::testing::Range(2, 9), ::testing::Range(0, 4));
+                          ::testing::Range(2, 10), ::testing::Range(0, 4));
 
 VP9_INSTANTIATE_TEST_CASE(DatarateTestVP9LargeOneBR,
                           ::testing::Values(::libvpx_test::kOnePassGood,
                                             ::libvpx_test::kRealTime),
-                          ::testing::Range(2, 9));
+                          ::testing::Range(2, 10));
 
-VP9_INSTANTIATE_TEST_CASE(DatarateTestVP9RealTime, ::testing::Range(5, 9));
+VP9_INSTANTIATE_TEST_CASE(DatarateTestVP9RealTime, ::testing::Range(5, 10));
 
 #if CONFIG_VP9_TEMPORAL_DENOISING
-VP9_INSTANTIATE_TEST_CASE(DatarateTestVP9LargeDenoiser, ::testing::Range(5, 9));
+VP9_INSTANTIATE_TEST_CASE(DatarateTestVP9LargeDenoiser,
+                          ::testing::Range(5, 10));
 #endif
 }  // namespace
diff --git a/test/vp9_ethread_test.cc b/test/vp9_ethread_test.cc
index 6b7e51211..44659904f 100644
--- a/test/vp9_ethread_test.cc
+++ b/test/vp9_ethread_test.cc
@@ -409,7 +409,7 @@ INSTANTIATE_TEST_CASE_P(
         ::testing::Values(::libvpx_test::kTwoPassGood,
                           ::libvpx_test::kOnePassGood,
                           ::libvpx_test::kRealTime),
-        ::testing::Range(3, 9),    // cpu_used
+        ::testing::Range(3, 10),   // cpu_used
         ::testing::Range(0, 3),    // tile_columns
         ::testing::Range(2, 5)));  // threads
 
diff --git a/vp9/encoder/vp9_pickmode.c b/vp9/encoder/vp9_pickmode.c
index f7d420bc7..f86c9f092 100644
--- a/vp9/encoder/vp9_pickmode.c
+++ b/vp9/encoder/vp9_pickmode.c
@@ -224,6 +224,14 @@ static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
   if (rv && search_subpel) {
     int subpel_force_stop = cpi->sf.mv.subpel_force_stop;
     if (use_base_mv && cpi->sf.base_mv_aggressive) subpel_force_stop = 2;
+    if (cpi->sf.mv.enable_adaptive_subpel_force_stop) {
+      int mv_thresh = cpi->sf.mv.adapt_subpel_force_stop.mv_thresh;
+      if (abs(tmp_mv->as_mv.row) >= mv_thresh ||
+          abs(tmp_mv->as_mv.col) >= mv_thresh)
+        subpel_force_stop = cpi->sf.mv.adapt_subpel_force_stop.force_stop_above;
+      else
+        subpel_force_stop = cpi->sf.mv.adapt_subpel_force_stop.force_stop_below;
+    }
     cpi->find_fractional_mv_step(
         x, &tmp_mv->as_mv, &ref_mv, cpi->common.allow_high_precision_mv,
         x->errorperbit, &cpi->fn_ptr[bsize], subpel_force_stop,
diff --git a/vp9/encoder/vp9_speed_features.c b/vp9/encoder/vp9_speed_features.c
index 90da68726..05ec3c612 100644
--- a/vp9/encoder/vp9_speed_features.c
+++ b/vp9/encoder/vp9_speed_features.c
@@ -661,6 +661,14 @@ static void set_rt_speed_feature_framesize_independent(
     sf->limit_newmv_early_exit = 0;
     sf->use_simple_block_yrd = 1;
   }
+
+  if (speed >= 9) {
+    sf->mv.enable_adaptive_subpel_force_stop = 1;
+    sf->mv.adapt_subpel_force_stop.mv_thresh = 2;
+    sf->mv.adapt_subpel_force_stop.force_stop_below = 1;
+    sf->mv.adapt_subpel_force_stop.force_stop_above = 2;
+  }
+
   if (sf->use_altref_onepass) {
     if (cpi->rc.is_src_frame_alt_ref && cm->frame_type != KEY_FRAME) {
       sf->partition_search_type = FIXED_PARTITION;
diff --git a/vp9/encoder/vp9_speed_features.h b/vp9/encoder/vp9_speed_features.h
index 946bf0545..8595e54ab 100644
--- a/vp9/encoder/vp9_speed_features.h
+++ b/vp9/encoder/vp9_speed_features.h
@@ -161,6 +161,17 @@ typedef enum {
   ONE_LOOP_REDUCED = 1
 } FAST_COEFF_UPDATE;
 
+typedef struct ADAPT_SUBPEL_FORCE_STOP {
+  // Threshold for full pixel motion vector;
+  int mv_thresh;
+
+  // subpel_force_stop if full pixel MV is below the threshold.
+  int force_stop_below;
+
+  // subpel_force_stop if full pixel MV is equal to or above the threshold.
+  int force_stop_above;
+} ADAPT_SUBPEL_FORCE_STOP;
+
 typedef struct MV_SPEED_FEATURES {
   // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
   SEARCH_METHODS search_method;
@@ -189,6 +200,11 @@ typedef struct MV_SPEED_FEATURES {
   // 3: Stop at full pixel.
   int subpel_force_stop;
 
+  // If it's enabled, different subpel_force_stop will be used for different MV.
+  int enable_adaptive_subpel_force_stop;
+
+  ADAPT_SUBPEL_FORCE_STOP adapt_subpel_force_stop;
+
   // This variable sets the step_param used in full pel motion search.
   int fullpel_search_step_param;
 } MV_SPEED_FEATURES;
diff --git a/vp9/vp9_cx_iface.c b/vp9/vp9_cx_iface.c
index 5eaa7a18a..d6c6ece91 100644
--- a/vp9/vp9_cx_iface.c
+++ b/vp9/vp9_cx_iface.c
@@ -248,7 +248,7 @@ static vpx_codec_err_t validate_config(vpx_codec_alg_priv_t *ctx,
   RANGE_CHECK(extra_cfg, row_mt, 0, 1);
   RANGE_CHECK(extra_cfg, motion_vector_unit_test, 0, 2);
   RANGE_CHECK(extra_cfg, enable_auto_alt_ref, 0, 2);
-  RANGE_CHECK(extra_cfg, cpu_used, -8, 8);
+  RANGE_CHECK(extra_cfg, cpu_used, -9, 9);
   RANGE_CHECK_HI(extra_cfg, noise_sensitivity, 6);
   RANGE_CHECK(extra_cfg, tile_columns, 0, 6);
   RANGE_CHECK(extra_cfg, tile_rows, 0, 2);