vp9: Fix some issues with denoiser and SVC.

Fix the update of the denoiser buffer when the base
spatial layer is a key frame. And allow for better/lower
QP on high spatial layers when their base layer is key frame.

Change-Id: I96b2426f1eaa43b8b8d4c31a68b0c6d68c3024a2

diff --git a/vp9/encoder/vp9_denoiser.c b/vp9/encoder/vp9_denoiser.c
index c16429c..c995a9d 100644 (file)
--- a/vp9/encoder/vp9_denoiser.c
+++ b/vp9/encoder/vp9_denoiser.c
@@ -413,15 +413,15 @@ static void swap_frame_buffer(YV12_BUFFER_CONFIG *const dest,
   src->y_buffer = tmp_buf;
 }
 
-void vp9_denoiser_update_frame_info(VP9_DENOISER *denoiser,
-                                    YV12_BUFFER_CONFIG src,
-                                    FRAME_TYPE frame_type,
-                                    int refresh_alt_ref_frame,
-                                    int refresh_golden_frame,
-                                    int refresh_last_frame, int resized) {
+void vp9_denoiser_update_frame_info(
+    VP9_DENOISER *denoiser, YV12_BUFFER_CONFIG src, FRAME_TYPE frame_type,
+    int refresh_alt_ref_frame, int refresh_golden_frame, int refresh_last_frame,
+    int resized, int svc_base_is_key) {
   // Copy source into denoised reference buffers on KEY_FRAME or
-  // if the just encoded frame was resized.
-  if (frame_type == KEY_FRAME || resized != 0 || denoiser->reset) {
+  // if the just encoded frame was resized. For SVC, copy source if the base
+  // spatial layer was key frame.
+  if (frame_type == KEY_FRAME || resized != 0 || denoiser->reset ||
+      svc_base_is_key) {
     int i;
     // Start at 1 so as not to overwrite the INTRA_FRAME
     for (i = 1; i < MAX_REF_FRAMES; ++i)

diff --git a/vp9/encoder/vp9_denoiser.h b/vp9/encoder/vp9_denoiser.h
index 0ec8622..ce9a596 100644 (file)
--- a/vp9/encoder/vp9_denoiser.h
+++ b/vp9/encoder/vp9_denoiser.h
@@ -59,12 +59,10 @@ typedef struct {
 
 struct VP9_COMP;
 
-void vp9_denoiser_update_frame_info(VP9_DENOISER *denoiser,
-                                    YV12_BUFFER_CONFIG src,
-                                    FRAME_TYPE frame_type,
-                                    int refresh_alt_ref_frame,
-                                    int refresh_golden_frame,
-                                    int refresh_last_frame, int resized);
+void vp9_denoiser_update_frame_info(
+    VP9_DENOISER *denoiser, YV12_BUFFER_CONFIG src, FRAME_TYPE frame_type,
+    int refresh_alt_ref_frame, int refresh_golden_frame, int refresh_last_frame,
+    int resized, int svc_base_is_key);
 
 void vp9_denoiser_denoise(struct VP9_COMP *cpi, MACROBLOCK *mb, int mi_row,
                           int mi_col, BLOCK_SIZE bs, PICK_MODE_CONTEXT *ctx,

diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index 0d51f9b..5a7351a 100644 (file)
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -2555,10 +2555,18 @@ void vp9_update_reference_frames(VP9_COMP *cpi) {
 #if CONFIG_VP9_TEMPORAL_DENOISING
   if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
       cpi->denoiser.denoising_level > kDenLowLow) {
+    int svc_base_is_key = 0;
+    if (cpi->use_svc) {
+      int layer = LAYER_IDS_TO_IDX(cpi->svc.spatial_layer_id,
+                                   cpi->svc.temporal_layer_id,
+                                   cpi->svc.number_temporal_layers);
+      LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer];
+      svc_base_is_key = lc->is_key_frame;
+    }
     vp9_denoiser_update_frame_info(
         &cpi->denoiser, *cpi->Source, cpi->common.frame_type,
         cpi->refresh_alt_ref_frame, cpi->refresh_golden_frame,
-        cpi->refresh_last_frame, cpi->resize_pending);
+        cpi->refresh_last_frame, cpi->resize_pending, svc_base_is_key);
   }
 #endif
   if (is_one_pass_cbr_svc(cpi)) {

diff --git a/vp9/encoder/vp9_noise_estimate.c b/vp9/encoder/vp9_noise_estimate.c
index a32e5ca..fd6a4ee 100644 (file)
--- a/vp9/encoder/vp9_noise_estimate.c
+++ b/vp9/encoder/vp9_noise_estimate.c
@@ -131,7 +131,8 @@ void vp9_update_noise_estimate(VP9_COMP *const cpi) {
     // Force noise estimation to 0 and denoiser off if content has high motion.
     ne->level = kLowLow;
 #if CONFIG_VP9_TEMPORAL_DENOISING
-    if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi))
+    if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
+        cpi->svc.current_superframe > 1)
       vp9_denoiser_set_noise_level(&cpi->denoiser, ne->level);
 #endif
     return;

diff --git a/vp9/encoder/vp9_ratectrl.c b/vp9/encoder/vp9_ratectrl.c
index 6932c0c..b95cd77 100644 (file)
--- a/vp9/encoder/vp9_ratectrl.c
+++ b/vp9/encoder/vp9_ratectrl.c
@@ -688,6 +688,17 @@ static int calc_active_worst_quality_one_pass_cbr(const VP9_COMP *cpi) {
                    ? VPXMIN(rc->avg_frame_qindex[INTER_FRAME],
                             rc->avg_frame_qindex[KEY_FRAME])
                    : rc->avg_frame_qindex[INTER_FRAME];
+  // For SVC if the current base spatial layer was key frame, use the QP from
+  // that base layer for ambient_qp.
+  if (cpi->use_svc && cpi->svc.spatial_layer_id > 0) {
+    int layer = LAYER_IDS_TO_IDX(0, cpi->svc.temporal_layer_id,
+                                 cpi->svc.number_temporal_layers);
+    const LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer];
+    if (lc->is_key_frame) {
+      const RATE_CONTROL *lrc = &lc->rc;
+      ambient_qp = VPXMIN(ambient_qp, lrc->last_q[KEY_FRAME]);
+    }
+  }
   active_worst_quality = VPXMIN(rc->worst_quality, ambient_qp * 5 >> 2);
   if (rc->buffer_level > rc->optimal_buffer_level) {
     // Adjust down.