Merge "Port "cost_coeff speed improvements" to vp9."
diff --git a/examples/vp8_multi_resolution_encoder.c b/examples/vp8_multi_resolution_encoder.c
index 0248ede..fc775ef 100644
--- a/examples/vp8_multi_resolution_encoder.c
+++ b/examples/vp8_multi_resolution_encoder.c
@@ -347,8 +347,7 @@
double psnr_totals[NUM_ENCODERS][4] = {{0,0}};
int psnr_count[NUM_ENCODERS] = {0};
- double cx_time = 0;
- struct timeval tv1, tv2, difftv;
+ int64_t cx_time = 0;
/* Set the required target bitrates for each resolution level.
* If target bitrate for highest-resolution level is set to 0,
@@ -582,6 +581,7 @@
while(frame_avail || got_data)
{
+ struct vpx_usec_timer timer;
vpx_codec_iter_t iter[NUM_ENCODERS]={NULL};
const vpx_codec_cx_pkt_t *pkt[NUM_ENCODERS];
@@ -636,18 +636,18 @@
vpx_codec_control(&codec[i], VP8E_SET_TEMPORAL_LAYER_ID, layer_id);
}
- gettimeofday(&tv1, NULL);
/* Encode each frame at multi-levels */
/* Note the flags must be set to 0 in the encode call if they are set
for each frame with the vpx_codec_control(), as done above. */
+ vpx_usec_timer_start(&timer);
if(vpx_codec_encode(&codec[0], frame_avail? &raw[0] : NULL,
frame_cnt, 1, 0, arg_deadline))
{
die_codec(&codec[0], "Failed to encode frame");
}
- gettimeofday(&tv2, NULL);
- timersub(&tv2, &tv1, &difftv);
- cx_time += (double)(difftv.tv_sec * 1000000 + difftv.tv_usec);
+ vpx_usec_timer_mark(&timer);
+ cx_time += vpx_usec_timer_elapsed(&timer);
+
for (i=NUM_ENCODERS-1; i>=0 ; i--)
{
got_data = 0;
@@ -686,8 +686,10 @@
frame_cnt++;
}
printf("\n");
- printf("FPS for encoding %d %f %f \n", frame_cnt, (float)cx_time / 1000000,
- 1000000 * (double)frame_cnt / (double)cx_time);
+ printf("Frame cnt and encoding time/FPS stats for encoding: %d %f %f \n",
+ frame_cnt,
+ 1000 * (float)cx_time / (double)(frame_cnt * 1000000),
+ 1000000 * (double)frame_cnt / (double)cx_time);
fclose(infile);
diff --git a/test/datarate_test.cc b/test/datarate_test.cc
index e94df20..d6b1b52 100644
--- a/test/datarate_test.cc
+++ b/test/datarate_test.cc
@@ -519,6 +519,9 @@
cfg_.rc_end_usage = VPX_CBR;
cfg_.rc_target_bitrate = 200;
cfg_.g_lag_in_frames = 0;
+ // TODO(marpan): Investigate datarate target failures with a smaller keyframe
+ // interval (128).
+ cfg_.kf_max_dist = 9999;
::libvpx_test::I420VideoSource video("hantro_collage_w352h288.yuv", 352, 288,
30, 1, 0, 140);
@@ -774,10 +777,6 @@
svc_params_.max_quantizers[i] = 63;
svc_params_.min_quantizers[i] = 0;
}
- svc_params_.scaling_factor_num[0] = 144;
- svc_params_.scaling_factor_den[0] = 288;
- svc_params_.scaling_factor_num[1] = 288;
- svc_params_.scaling_factor_den[1] = 288;
encoder->Control(VP9E_SET_SVC, 1);
encoder->Control(VP9E_SET_SVC_PARAMETERS, &svc_params_);
encoder->Control(VP8E_SET_CPUUSED, speed_setting_);
@@ -814,8 +813,6 @@
if (bits_total_) {
const double file_size_in_kb = bits_total_ / 1000.; // bits per kilobit
duration_ = (last_pts_ + 1) * timebase_;
- effective_datarate_ = (bits_total_ - bits_in_last_frame_) / 1000.0
- / (cfg_.rc_buf_initial_sz / 1000.0 + duration_);
file_datarate_ = file_size_in_kb / duration_;
}
}
@@ -839,7 +836,6 @@
int64_t bits_total_;
double duration_;
double file_datarate_;
- double effective_datarate_;
size_t bits_in_last_frame_;
vpx_svc_extra_cfg_t svc_params_;
int speed_setting_;
@@ -884,7 +880,7 @@
// Check basic rate targeting for 1 pass CBR SVC: 2 spatial layers and
// 3 temporal layers. Run CIF clip with 1 thread.
-TEST_P(DatarateOnePassCbrSvc, OnePassCbrSvc) {
+TEST_P(DatarateOnePassCbrSvc, OnePassCbrSvc2SpatialLayers) {
cfg_.rc_buf_initial_sz = 500;
cfg_.rc_buf_optimal_sz = 500;
cfg_.rc_buf_sz = 1000;
@@ -918,7 +914,7 @@
assign_layer_bitrates(&cfg_, &svc_params_, cfg_.ss_number_layers,
cfg_.ts_number_layers, cfg_.temporal_layering_mode);
ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
- ASSERT_GE(cfg_.rc_target_bitrate, effective_datarate_ * 0.85)
+ ASSERT_GE(cfg_.rc_target_bitrate, file_datarate_ * 0.85)
<< " The datarate for the file exceeds the target by too much!";
ASSERT_LE(cfg_.rc_target_bitrate, file_datarate_ * 1.15)
<< " The datarate for the file is lower than the target by too much!";
@@ -926,9 +922,49 @@
}
}
+// Check basic rate targeting for 1 pass CBR SVC: 3 spatial layers and
+// 3 temporal layers. Run CIF clip with 1 thread.
+TEST_P(DatarateOnePassCbrSvc, OnePassCbrSvc3SpatialLayers) {
+ cfg_.rc_buf_initial_sz = 500;
+ cfg_.rc_buf_optimal_sz = 500;
+ cfg_.rc_buf_sz = 1000;
+ cfg_.rc_min_quantizer = 0;
+ cfg_.rc_max_quantizer = 63;
+ cfg_.rc_end_usage = VPX_CBR;
+ cfg_.g_lag_in_frames = 0;
+ cfg_.ss_number_layers = 3;
+ cfg_.ts_number_layers = 3;
+ cfg_.ts_rate_decimator[0] = 4;
+ cfg_.ts_rate_decimator[1] = 2;
+ cfg_.ts_rate_decimator[2] = 1;
+ cfg_.g_error_resilient = 1;
+ cfg_.g_threads = 1;
+ cfg_.temporal_layering_mode = 3;
+ svc_params_.scaling_factor_num[0] = 72;
+ svc_params_.scaling_factor_den[0] = 288;
+ svc_params_.scaling_factor_num[1] = 144;
+ svc_params_.scaling_factor_den[1] = 288;
+ svc_params_.scaling_factor_num[2] = 288;
+ svc_params_.scaling_factor_den[2] = 288;
+ cfg_.rc_dropframe_thresh = 10;
+ cfg_.kf_max_dist = 9999;
+ ::libvpx_test::I420VideoSource video("niklas_1280_720_30.y4m", 1280, 720,
+ 30, 1, 0, 300);
+ cfg_.rc_target_bitrate = 800;
+ ResetModel();
+ assign_layer_bitrates(&cfg_, &svc_params_, cfg_.ss_number_layers,
+ cfg_.ts_number_layers, cfg_.temporal_layering_mode);
+ ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+ ASSERT_GE(cfg_.rc_target_bitrate, file_datarate_ * 0.85)
+ << " The datarate for the file exceeds the target by too much!";
+ ASSERT_LE(cfg_.rc_target_bitrate, file_datarate_ * 1.22)
+ << " The datarate for the file is lower than the target by too much!";
+ EXPECT_EQ(GetMismatchFrames(), (unsigned int) 0);
+}
+
// Check basic rate targeting for 1 pass CBR SVC: 2 spatial layers and
// 3 temporal layers. Run HD clip with 4 threads.
-TEST_P(DatarateOnePassCbrSvc, OnePassCbrSvc4threads) {
+TEST_P(DatarateOnePassCbrSvc, OnePassCbrSvc2SpatialLayers4threads) {
cfg_.rc_buf_initial_sz = 500;
cfg_.rc_buf_optimal_sz = 500;
cfg_.rc_buf_sz = 1000;
@@ -949,6 +985,7 @@
svc_params_.scaling_factor_num[1] = 288;
svc_params_.scaling_factor_den[1] = 288;
cfg_.rc_dropframe_thresh = 10;
+ cfg_.kf_max_dist = 9999;
::libvpx_test::I420VideoSource video("niklas_1280_720_30.y4m", 1280, 720,
30, 1, 0, 300);
cfg_.rc_target_bitrate = 800;
@@ -956,19 +993,59 @@
assign_layer_bitrates(&cfg_, &svc_params_, cfg_.ss_number_layers,
cfg_.ts_number_layers, cfg_.temporal_layering_mode);
ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
- ASSERT_GE(cfg_.rc_target_bitrate, effective_datarate_ * 0.85)
+ ASSERT_GE(cfg_.rc_target_bitrate, file_datarate_ * 0.85)
<< " The datarate for the file exceeds the target by too much!";
ASSERT_LE(cfg_.rc_target_bitrate, file_datarate_ * 1.15)
<< " The datarate for the file is lower than the target by too much!";
EXPECT_EQ(GetMismatchFrames(), (unsigned int) 0);
}
+// Check basic rate targeting for 1 pass CBR SVC: 3 spatial layers and
+// 3 temporal layers. Run HD clip with 4 threads.
+TEST_P(DatarateOnePassCbrSvc, OnePassCbrSvc3SpatialLayers4threads) {
+ cfg_.rc_buf_initial_sz = 500;
+ cfg_.rc_buf_optimal_sz = 500;
+ cfg_.rc_buf_sz = 1000;
+ cfg_.rc_min_quantizer = 0;
+ cfg_.rc_max_quantizer = 63;
+ cfg_.rc_end_usage = VPX_CBR;
+ cfg_.g_lag_in_frames = 0;
+ cfg_.ss_number_layers = 3;
+ cfg_.ts_number_layers = 3;
+ cfg_.ts_rate_decimator[0] = 4;
+ cfg_.ts_rate_decimator[1] = 2;
+ cfg_.ts_rate_decimator[2] = 1;
+ cfg_.g_error_resilient = 1;
+ cfg_.g_threads = 4;
+ cfg_.temporal_layering_mode = 3;
+ svc_params_.scaling_factor_num[0] = 72;
+ svc_params_.scaling_factor_den[0] = 288;
+ svc_params_.scaling_factor_num[1] = 144;
+ svc_params_.scaling_factor_den[1] = 288;
+ svc_params_.scaling_factor_num[2] = 288;
+ svc_params_.scaling_factor_den[2] = 288;
+ cfg_.rc_dropframe_thresh = 10;
+ cfg_.kf_max_dist = 9999;
+ ::libvpx_test::I420VideoSource video("niklas_1280_720_30.y4m", 1280, 720,
+ 30, 1, 0, 300);
+ cfg_.rc_target_bitrate = 800;
+ ResetModel();
+ assign_layer_bitrates(&cfg_, &svc_params_, cfg_.ss_number_layers,
+ cfg_.ts_number_layers, cfg_.temporal_layering_mode);
+ ASSERT_NO_FATAL_FAILURE(RunLoop(&video));
+ ASSERT_GE(cfg_.rc_target_bitrate, file_datarate_ * 0.85)
+ << " The datarate for the file exceeds the target by too much!";
+ ASSERT_LE(cfg_.rc_target_bitrate, file_datarate_ * 1.22)
+ << " The datarate for the file is lower than the target by too much!";
+ EXPECT_EQ(GetMismatchFrames(), (unsigned int) 0);
+}
+
VP8_INSTANTIATE_TEST_CASE(DatarateTestLarge, ALL_TEST_MODES);
VP9_INSTANTIATE_TEST_CASE(DatarateTestVP9Large,
::testing::Values(::libvpx_test::kOnePassGood,
::libvpx_test::kRealTime),
- ::testing::Range(2, 7));
+ ::testing::Range(2, 9));
VP9_INSTANTIATE_TEST_CASE(DatarateOnePassCbrSvc,
::testing::Values(::libvpx_test::kRealTime),
- ::testing::Range(5, 8));
+ ::testing::Range(5, 9));
} // namespace
diff --git a/vp8/encoder/denoising.c b/vp8/encoder/denoising.c
index 113865f..5ae44e8 100644
--- a/vp8/encoder/denoising.c
+++ b/vp8/encoder/denoising.c
@@ -23,7 +23,7 @@
*/
static const unsigned int SSE_DIFF_THRESHOLD = 16 * 16 * 20;
static const unsigned int SSE_THRESHOLD = 16 * 16 * 40;
-static const unsigned int SSE_THRESHOLD_HIGH = 16 * 16 * 60;
+static const unsigned int SSE_THRESHOLD_HIGH = 16 * 16 * 80;
/*
* The filter function was modified to reduce the computational complexity.
diff --git a/vp8/encoder/denoising.h b/vp8/encoder/denoising.h
index 9a379a6..f13d52f 100644
--- a/vp8/encoder/denoising.h
+++ b/vp8/encoder/denoising.h
@@ -18,8 +18,8 @@
extern "C" {
#endif
-#define SUM_DIFF_THRESHOLD (16 * 16 * 2)
-#define SUM_DIFF_THRESHOLD_HIGH (600) // ~(16 * 16 * 1.5)
+#define SUM_DIFF_THRESHOLD 384
+#define SUM_DIFF_THRESHOLD_HIGH 512
#define MOTION_MAGNITUDE_THRESHOLD (8*3)
#define SUM_DIFF_THRESHOLD_UV (96) // (8 * 8 * 1.5)
diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c
index 4109c19..8953f0f 100644
--- a/vp9/encoder/vp9_encodeframe.c
+++ b/vp9/encoder/vp9_encodeframe.c
@@ -770,6 +770,7 @@
x->pred_mv[LAST_FRAME] = mi->mv[0].as_mv;
}
+ set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
vp9_build_inter_predictors_sb(xd, mi_row, mi_col, BLOCK_64X64);
// Check if most of the superblock is skin content, and if so, force split
@@ -2439,7 +2440,8 @@
PARTITION_CONTEXT sl[8], sa[8];
TOKENEXTRA *tp_orig = *tp;
PICK_MODE_CONTEXT *ctx = &pc_tree->none;
- int i, pl;
+ int i;
+ const int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
BLOCK_SIZE subsize;
RD_COST this_rdc, sum_rdc, best_rdc;
int do_split = bsize >= BLOCK_8X8;
@@ -2587,7 +2589,6 @@
&this_rdc, bsize, ctx, best_rdc.rdcost);
if (this_rdc.rate != INT_MAX) {
if (bsize >= BLOCK_8X8) {
- pl = partition_plane_context(xd, mi_row, mi_col, bsize);
this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
this_rdc.rate, this_rdc.dist);
@@ -2706,7 +2707,6 @@
}
if (sum_rdc.rdcost < best_rdc.rdcost && i == 4) {
- pl = partition_plane_context(xd, mi_row, mi_col, bsize);
sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
sum_rdc.rate, sum_rdc.dist);
@@ -2772,7 +2772,6 @@
}
if (sum_rdc.rdcost < best_rdc.rdcost) {
- pl = partition_plane_context(xd, mi_row, mi_col, bsize);
sum_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
if (sum_rdc.rdcost < best_rdc.rdcost) {
@@ -2824,7 +2823,6 @@
}
if (sum_rdc.rdcost < best_rdc.rdcost) {
- pl = partition_plane_context(xd, mi_row, mi_col, bsize);
sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
sum_rdc.rate, sum_rdc.dist);
@@ -4271,13 +4269,9 @@
VP9_COMMON *const cm = &cpi->common;
MACROBLOCK *const x = &td->mb;
MACROBLOCKD *const xd = &x->e_mbd;
- MODE_INFO **mi_8x8 = xd->mi;
- MODE_INFO *mi = mi_8x8[0];
+ MODE_INFO *mi = xd->mi[0];
const int seg_skip = segfeature_active(&cm->seg, mi->segment_id,
SEG_LVL_SKIP);
- const int mis = cm->mi_stride;
- const int mi_width = num_8x8_blocks_wide_lookup[bsize];
- const int mi_height = num_8x8_blocks_high_lookup[bsize];
x->skip_recode = !x->select_tx_size && mi->sb_type >= BLOCK_8X8 &&
cpi->oxcf.aq_mode != COMPLEXITY_AQ &&
@@ -4333,20 +4327,14 @@
++get_tx_counts(max_txsize_lookup[bsize], get_tx_size_context(xd),
&td->counts->tx)[mi->tx_size];
} else {
- int x, y;
- TX_SIZE tx_size;
// The new intra coding scheme requires no change of transform size
if (is_inter_block(mi)) {
- tx_size = VPXMIN(tx_mode_to_biggest_tx_size[cm->tx_mode],
- max_txsize_lookup[bsize]);
+ mi->tx_size = VPXMIN(tx_mode_to_biggest_tx_size[cm->tx_mode],
+ max_txsize_lookup[bsize]);
} else {
- tx_size = (bsize >= BLOCK_8X8) ? mi->tx_size : TX_4X4;
+ mi->tx_size = (bsize >= BLOCK_8X8) ? mi->tx_size : TX_4X4;
}
- for (y = 0; y < mi_height; y++)
- for (x = 0; x < mi_width; x++)
- if (mi_col + x < cm->mi_cols && mi_row + y < cm->mi_rows)
- mi_8x8[mis * y + x]->tx_size = tx_size;
}
++td->counts->tx.tx_totals[mi->tx_size];
++td->counts->tx.tx_totals[get_uv_tx_size(mi, &xd->plane[1])];
diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index f8f681a..6d6915c 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -410,6 +410,9 @@
memset(&cpi->svc.scaled_frames[0], 0,
MAX_LAG_BUFFERS * sizeof(cpi->svc.scaled_frames[0]));
+ vpx_free_frame_buffer(&cpi->svc.scaled_temp);
+ memset(&cpi->svc.scaled_temp, 0, sizeof(cpi->svc.scaled_temp));
+
vpx_free_frame_buffer(&cpi->svc.empty_frame.img);
memset(&cpi->svc.empty_frame, 0, sizeof(cpi->svc.empty_frame));
@@ -3357,11 +3360,22 @@
vpx_clear_system_state();
set_frame_size(cpi);
- cpi->Source = vp9_scale_if_required(cm,
- cpi->un_scaled_source,
- &cpi->scaled_source,
- (cpi->oxcf.pass == 0));
+ if (is_one_pass_cbr_svc(cpi) &&
+ cpi->un_scaled_source->y_width == cm->width << 2 &&
+ cpi->un_scaled_source->y_height == cm->height << 2 &&
+ cpi->svc.scaled_temp.y_width == cm->width << 1 &&
+ cpi->svc.scaled_temp.y_height == cm->height << 1) {
+ cpi->Source = vp9_svc_twostage_scale(cm,
+ cpi->un_scaled_source,
+ &cpi->scaled_source,
+ &cpi->svc.scaled_temp);
+ } else {
+ cpi->Source = vp9_scale_if_required(cm,
+ cpi->un_scaled_source,
+ &cpi->scaled_source,
+ (cpi->oxcf.pass == 0));
+ }
// Avoid scaling last_source unless its needed.
// Last source is needed if vp9_avg_source_sad() is used, or if
// partition_search_type == SOURCE_VAR_BASED_PARTITION, or if noise
@@ -3387,12 +3401,10 @@
cpi->oxcf.rc_mode == VPX_VBR))
vp9_avg_source_sad(cpi);
- // TODO(wonkap/marpan): For 1 pass SVC, since only ZERMOV is allowed for
- // upsampled reference frame (i.e, svc->force_zero_mode_spatial_ref = 0),
- // we should be able to avoid this frame-level upsampling.
- // Keeping it for now as there is an asan error in the multi-threaded SVC
- // rate control test if this upsampling is removed.
- if (frame_is_intra_only(cm) == 0) {
+ // For 1 pass SVC, since only ZEROMV is allowed for upsampled reference
+ // frame (i.e, svc->force_zero_mode_spatial_ref = 0), we can avoid this
+ // frame-level upsampling.
+ if (frame_is_intra_only(cm) == 0 && !is_one_pass_cbr_svc(cpi)) {
vp9_scale_references(cpi);
}
@@ -3782,6 +3794,25 @@
}
}
+YV12_BUFFER_CONFIG *vp9_svc_twostage_scale(VP9_COMMON *cm,
+ YV12_BUFFER_CONFIG *unscaled,
+ YV12_BUFFER_CONFIG *scaled,
+ YV12_BUFFER_CONFIG *scaled_temp) {
+ if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
+ cm->mi_rows * MI_SIZE != unscaled->y_height) {
+#if CONFIG_VP9_HIGHBITDEPTH
+ scale_and_extend_frame(unscaled, scaled_temp, (int)cm->bit_depth);
+ scale_and_extend_frame(scaled_temp, scaled, (int)cm->bit_depth);
+#else
+ vp9_scale_and_extend_frame(unscaled, scaled_temp);
+ vp9_scale_and_extend_frame(scaled_temp, scaled);
+#endif // CONFIG_VP9_HIGHBITDEPTH
+ return scaled;
+ } else {
+ return unscaled;
+ }
+}
+
YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
YV12_BUFFER_CONFIG *unscaled,
YV12_BUFFER_CONFIG *scaled,
diff --git a/vp9/encoder/vp9_encoder.h b/vp9/encoder/vp9_encoder.h
index c486ac2..2def941 100644
--- a/vp9/encoder/vp9_encoder.h
+++ b/vp9/encoder/vp9_encoder.h
@@ -623,6 +623,11 @@
void vp9_set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv);
+YV12_BUFFER_CONFIG *vp9_svc_twostage_scale(VP9_COMMON *cm,
+ YV12_BUFFER_CONFIG *unscaled,
+ YV12_BUFFER_CONFIG *scaled,
+ YV12_BUFFER_CONFIG *scaled_temp);
+
YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
YV12_BUFFER_CONFIG *unscaled,
YV12_BUFFER_CONFIG *scaled,
diff --git a/vp9/encoder/vp9_svc_layercontext.c b/vp9/encoder/vp9_svc_layercontext.c
index 1d56154..79e5049 100644
--- a/vp9/encoder/vp9_svc_layercontext.c
+++ b/vp9/encoder/vp9_svc_layercontext.c
@@ -43,6 +43,26 @@
cpi->svc.ext_alt_fb_idx[sl] = 2;
}
+ // For 1 pass cbr: allocate scaled_frame that may be used as an intermediate
+ // buffer for a 2 stage down-sampling: two stages of 1:2 down-sampling for a
+ // target of 1/4x1/4.
+ if (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_CBR) {
+ if (vpx_realloc_frame_buffer(&cpi->svc.scaled_temp,
+ cpi->common.width >> 1,
+ cpi->common.height >> 1,
+ cpi->common.subsampling_x,
+ cpi->common.subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+ cpi->common.use_highbitdepth,
+#endif
+ VP9_ENC_BORDER_IN_PIXELS,
+ cpi->common.byte_alignment,
+ NULL, NULL, NULL))
+ vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
+ "Failed to allocate scaled_frame for svc ");
+ }
+
+
if (cpi->oxcf.error_resilient_mode == 0 && cpi->oxcf.pass == 2) {
if (vpx_realloc_frame_buffer(&cpi->svc.empty_frame.img,
SMALL_FRAME_WIDTH, SMALL_FRAME_HEIGHT,
diff --git a/vp9/encoder/vp9_svc_layercontext.h b/vp9/encoder/vp9_svc_layercontext.h
index 4e18640..f1b8556 100644
--- a/vp9/encoder/vp9_svc_layercontext.h
+++ b/vp9/encoder/vp9_svc_layercontext.h
@@ -70,6 +70,8 @@
// Store scaled source frames to be used for temporal filter to generate
// a alt ref frame.
YV12_BUFFER_CONFIG scaled_frames[MAX_LAG_BUFFERS];
+ // Temp buffer used for 2-stage down-sampling, for real-time mode.
+ YV12_BUFFER_CONFIG scaled_temp;
// Layer context used for rate control in one pass temporal CBR mode or
// two pass spatial mode.
diff --git a/vp9/encoder/vp9_temporal_filter.c b/vp9/encoder/vp9_temporal_filter.c
index 82f566b..ebe28b8 100644
--- a/vp9/encoder/vp9_temporal_filter.c
+++ b/vp9/encoder/vp9_temporal_filter.c
@@ -45,8 +45,7 @@
int x, int y) {
const int which_mv = 0;
const MV mv = { mv_row, mv_col };
- const InterpKernel *const kernel =
- vp9_filter_kernels[xd->mi[0]->interp_filter];
+ const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP_SHARP];
enum mv_precision mv_precision_uv;
int uv_stride;
@@ -86,6 +85,7 @@
return;
}
#endif // CONFIG_VP9_HIGHBITDEPTH
+ (void)xd;
vp9_build_inter_predictor(y_mb_ptr, stride,
&pred[0], 16,
&mv,
diff --git a/vpx_dsp/x86/convolve.h b/vpx_dsp/x86/convolve.h
index b6fbfcf..7e43eb7 100644
--- a/vpx_dsp/x86/convolve.h
+++ b/vpx_dsp/x86/convolve.h
@@ -33,7 +33,7 @@
int w, int h) { \
assert(filter[3] != 128); \
assert(step_q4 == 16); \
- if (filter[0] || filter[1] || filter[2]) { \
+ if (filter[0] | filter[1] | filter[2]) { \
while (w >= 16) { \
vpx_filter_block1d16_##dir##8_##avg##opt(src_start, \
src_stride, \
@@ -45,27 +45,20 @@
dst += 16; \
w -= 16; \
} \
- while (w >= 8) { \
+ if (w == 8) { \
vpx_filter_block1d8_##dir##8_##avg##opt(src_start, \
src_stride, \
dst, \
dst_stride, \
h, \
filter); \
- src += 8; \
- dst += 8; \
- w -= 8; \
- } \
- while (w >= 4) { \
+ } else if (w == 4) { \
vpx_filter_block1d4_##dir##8_##avg##opt(src_start, \
src_stride, \
dst, \
dst_stride, \
h, \
filter); \
- src += 4; \
- dst += 4; \
- w -= 4; \
} \
} else { \
while (w >= 16) { \
@@ -79,27 +72,20 @@
dst += 16; \
w -= 16; \
} \
- while (w >= 8) { \
+ if (w == 8) { \
vpx_filter_block1d8_##dir##2_##avg##opt(src, \
src_stride, \
dst, \
dst_stride, \
h, \
filter); \
- src += 8; \
- dst += 8; \
- w -= 8; \
- } \
- while (w >= 4) { \
+ } else if (w == 4) { \
vpx_filter_block1d4_##dir##2_##avg##opt(src, \
src_stride, \
dst, \
dst_stride, \
h, \
filter); \
- src += 4; \
- dst += 4; \
- w -= 4; \
} \
} \
}
@@ -116,8 +102,7 @@
assert(h <= 64); \
assert(x_step_q4 == 16); \
assert(y_step_q4 == 16); \
- if (filter_x[0] || filter_x[1] || filter_x[2]|| \
- filter_y[0] || filter_y[1] || filter_y[2]) { \
+ if (filter_x[0] | filter_x[1] | filter_x[2]) { \
DECLARE_ALIGNED(16, uint8_t, fdata2[64 * 71]); \
vpx_convolve8_horiz_##opt(src - 3 * src_stride, src_stride, fdata2, 64, \
filter_x, x_step_q4, filter_y, y_step_q4, \
@@ -161,7 +146,7 @@
if (step_q4 == 16 && filter[3] != 128) { \
uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
- if (filter[0] || filter[1] || filter[2]) { \
+ if (filter[0] | filter[1] | filter[2]) { \
while (w >= 16) { \
vpx_highbd_filter_block1d16_##dir##8_##avg##opt(src_start, \
src_stride, \
@@ -253,8 +238,7 @@
assert(w <= 64); \
assert(h <= 64); \
if (x_step_q4 == 16 && y_step_q4 == 16) { \
- if (filter_x[0] || filter_x[1] || filter_x[2] || filter_x[3] == 128 || \
- filter_y[0] || filter_y[1] || filter_y[2] || filter_y[3] == 128) { \
+ if ((filter_x[0] | filter_x[1] | filter_x[2]) || filter_x[3] == 128) { \
DECLARE_ALIGNED(16, uint16_t, fdata2[64 * 71]); \
vpx_highbd_convolve8_horiz_##opt(src - 3 * src_stride, src_stride, \
CONVERT_TO_BYTEPTR(fdata2), 64, \
diff --git a/vpx_dsp/x86/vpx_subpixel_8t_ssse3.asm b/vpx_dsp/x86/vpx_subpixel_8t_ssse3.asm
index b0e99ed..d2cb8ea 100644
--- a/vpx_dsp/x86/vpx_subpixel_8t_ssse3.asm
+++ b/vpx_dsp/x86/vpx_subpixel_8t_ssse3.asm
@@ -16,6 +16,11 @@
; %define USE_PMULHRSW
; NOTE: pmulhrsw has a latency of 5 cycles. Tests showed a performance loss
; when using this instruction.
+;
+; The add order below (based on ffvp9) must be followed to prevent outranges.
+; x = k0k1 + k4k5
+; y = k2k3 + k6k7
+; z = signed SAT(x + y)
SECTION .text
%if ARCH_X86_64
@@ -77,17 +82,12 @@
pmaddubsw %2, k0k1k4k5
pmaddubsw m3, k2k3k6k7
-
- mova m4, %2
- mova m5, m3
- psrldq %2, 8
- psrldq m3, 8
- mova m6, m5
-
- paddsw m4, m3
- pmaxsw m5, %2
- pminsw %2, m6
+ mova m4, %2 ;k0k1
+ mova m5, m3 ;k2k3
+ psrldq %2, 8 ;k4k5
+ psrldq m3, 8 ;k6k7
paddsw %2, m4
+ paddsw m5, m3
paddsw %2, m5
paddsw %2, krd
psraw %2, 7
@@ -157,27 +157,20 @@
pmaddubsw m7, k0k1k4k5
palignr m3, m2, 5
pmaddubsw m3, k2k3k6k7
- mova m0, m4
- mova m5, m1
- mova m2, m7
- psrldq m4, 8
- psrldq m1, 8
- mova m6, m5
- paddsw m0, m1
- mova m1, m3
- psrldq m7, 8
- psrldq m3, 8
- paddsw m2, m3
- mova m3, m1
- pmaxsw m5, m4
- pminsw m4, m6
+ mova m0, m4 ;k0k1
+ mova m5, m1 ;k2k3
+ mova m2, m7 ;k0k1 upper
+ psrldq m4, 8 ;k4k5
+ psrldq m1, 8 ;k6k7
paddsw m4, m0
- paddsw m4, m5
- pmaxsw m1, m7
- pminsw m7, m3
+ paddsw m5, m1
+ mova m1, m3 ;k2k3 upper
+ psrldq m7, 8 ;k4k5 upper
+ psrldq m3, 8 ;k6k7 upper
paddsw m7, m2
+ paddsw m4, m5
+ paddsw m1, m3
paddsw m7, m1
-
paddsw m4, krd
psraw m4, 7
packuswb m4, m4
@@ -240,16 +233,13 @@
pmaddubsw %3, k2k3
pmaddubsw %4, k4k5
pmaddubsw %5, k6k7
-
+ paddsw %2, %4
+ paddsw %5, %3
paddsw %2, %5
- mova %1, %3
- pminsw %3, %4
- pmaxsw %1, %4
- paddsw %2, %3
- paddsw %1, %2
- paddsw %1, krd
- psraw %1, 7
- packuswb %1, %1
+ paddsw %2, krd
+ psraw %2, 7
+ packuswb %2, %2
+ SWAP %1, %2
%endm
;-------------------------------------------------------------------------------
@@ -293,39 +283,33 @@
pmaddubsw m3, k4k5
palignr m7, m4, 13
- paddsw m1, m5
- mova m5, m6
- mova m0, m2
- palignr m5, m4, 5
- pminsw m2, m3
- pmaddubsw m7, k6k7
- pmaxsw m3, m0
- paddsw m1, m2
mova m0, m6
- palignr m6, m4, 1
- pmaddubsw m5, k2k3
+ palignr m0, m4, 5
+ pmaddubsw m7, k6k7
paddsw m1, m3
+ paddsw m2, m5
+ paddsw m1, m2
+ mova m5, m6
+ palignr m6, m4, 1
+ pmaddubsw m0, k2k3
pmaddubsw m6, k0k1
- palignr m0, m4, 9
+ palignr m5, m4, 9
paddsw m1, krd
- pmaddubsw m0, k4k5
- mova m4, m5
+ pmaddubsw m5, k4k5
psraw m1, 7
- pminsw m5, m0
- paddsw m6, m7
+ paddsw m0, m7
+%ifidn %1, h8_avg
+ movh m7, [dstq]
+ movh m2, [dstq + dstrideq]
+%endif
packuswb m1, m1
-
paddsw m6, m5
- pmaxsw m0, m4
paddsw m6, m0
paddsw m6, krd
psraw m6, 7
packuswb m6, m6
-
%ifidn %1, h8_avg
- movh m0, [dstq]
- movh m2, [dstq + dstrideq]
- pavgb m1, m0
+ pavgb m1, m7
pavgb m6, m2
%endif
movh [dstq], m1
@@ -388,7 +372,7 @@
pmaddubsw m1, k2k3
palignr m2, m7, 9
pmaddubsw m2, k4k5
- paddsw m0, m3
+ paddsw m1, m3
mova m3, m4
punpckhbw m4, m4
mova m5, m4
@@ -403,17 +387,13 @@
pmaddubsw m6, k4k5
palignr m7, m3, 13
pmaddubsw m7, k6k7
-
- mova m3, m1
- pmaxsw m1, m2
- pminsw m2, m3
paddsw m0, m2
paddsw m0, m1
- paddsw m4, m7
- mova m7, m5
- pmaxsw m5, m6
- pminsw m6, m7
+%ifidn %1, h8_avg
+ mova m1, [dstq]
+%endif
paddsw m4, m6
+ paddsw m5, m7
paddsw m4, m5
paddsw m0, krd
paddsw m4, krd
@@ -421,7 +401,6 @@
psraw m4, 7
packuswb m0, m4
%ifidn %1, h8_avg
- mova m1, [dstq]
pavgb m0, m1
%endif
lea srcq, [srcq + sstrideq]
@@ -488,27 +467,21 @@
movx m7, [src1q + sstride6q ] ;H
punpcklbw m6, m7 ;G H
pmaddubsw m6, k6k7
- mova tmp, m2
pmaddubsw m3, k2k3
pmaddubsw m1, k0k1
- pmaxsw m2, m4
- paddsw m0, m6
+ paddsw m0, m4
+ paddsw m2, m6
movx m6, [srcq + sstrideq * 8 ] ;H next iter
punpcklbw m7, m6
pmaddubsw m7, k6k7
- pminsw m4, tmp
- paddsw m0, m4
- mova m4, m3
paddsw m0, m2
- pminsw m3, m5
- pmaxsw m5, m4
paddsw m0, krd
psraw m0, 7
- paddsw m1, m7
+ paddsw m1, m5
packuswb m0, m0
+ paddsw m3, m7
paddsw m1, m3
- paddsw m1, m5
paddsw m1, krd
psraw m1, 7
lea srcq, [srcq + sstrideq * 2 ]
@@ -550,10 +523,7 @@
punpcklbw m4, m5 ;E F
pmaddubsw m2, k2k3
pmaddubsw m4, k4k5
- paddsw m0, m6
- mova m1, m2
- pmaxsw m2, m4
- pminsw m4, m1
+ paddsw m2, m6
paddsw m0, m4
paddsw m0, m2
paddsw m0, krd
@@ -572,7 +542,6 @@
%macro SUBPIX_VFILTER16 1
cglobal filter_block1d16_%1, 6, 6+(ARCH_X86_64*3), 14, LOCAL_VARS_SIZE, \
src, sstride, dst, dstride, height, filter
-
mova m4, [filterq]
SETUP_LOCAL_VARS
%if ARCH_X86_64
@@ -611,12 +580,9 @@
punpcklbw m3, m5 ;A B
movh m7, [srcq + sstrideq * 2 + 8] ;C
pmaddubsw m6, k6k7
- mova m1, m2
movh m5, [src1q + sstrideq * 2 + 8] ;D
- pmaxsw m2, m4
punpcklbw m7, m5 ;C D
- pminsw m4, m1
- paddsw m0, m6
+ paddsw m2, m6
pmaddubsw m3, k0k1
movh m1, [srcq + sstrideq * 4 + 8] ;E
paddsw m0, m4
@@ -630,30 +596,24 @@
movh m5, [src1q + sstride6q + 8] ;H
psraw m0, 7
punpcklbw m2, m5 ;G H
- packuswb m0, m0
pmaddubsw m2, k6k7
%ifidn %1, v8_avg
- movh m4, [dstq]
- pavgb m0, m4
+ mova m4, [dstq]
%endif
movh [dstq], m0
- mova m6, m7
- pmaxsw m7, m1
- pminsw m1, m6
- paddsw m3, m2
+ paddsw m7, m2
paddsw m3, m1
paddsw m3, m7
paddsw m3, krd
psraw m3, 7
- packuswb m3, m3
+ packuswb m0, m3
add srcq, sstrideq
add src1q, sstrideq
%ifidn %1, v8_avg
- movh m1, [dstq + 8]
- pavgb m3, m1
+ pavgb m0, m4
%endif
- movh [dstq + 8], m3
+ mova [dstq], m0
add dstq, dst_stride
dec heightd
jnz .loop
