From: Minghai Shang <minghai@google.com>
Date: Tue, 14 Oct 2014 23:25:03 +0000 (-0700)
Subject: [spatial svc]Another workaround to avoid using prev_mi
X-Git-Tag: v1.4.0~602^2
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=68b550f5510a567583d9cdd9f60f44a280e611a7;p=libvpx

[spatial svc]Another workaround to avoid using prev_mi

We encode a empty invisible frame in front of the base layer frame to
avoid using prev_mi. Since there's a restriction for reference frame
scaling factor, we have to make it smaller and smaller gradually until
its size is 16x16.

Change remerged.

Change-Id: I9efab38bba7da86e056fbe8f663e711c5df38449
---

diff --git a/test/svc_test.cc b/test/svc_test.cc
index fdde702d2..67e83e395 100644
--- a/test/svc_test.cc
+++ b/test/svc_test.cc
@@ -225,10 +225,9 @@ class SvcTest : public ::testing::Test {
     EXPECT_EQ(received_frames, n);
   }
 
-  void DropLayersAndMakeItVP9Comaptible(struct vpx_fixed_buf *const inputs,
-                                        const int num_super_frames,
-                                        const int remained_spatial_layers,
-                                        const bool is_multiple_frame_contexts) {
+  void DropEnhancementLayers(struct vpx_fixed_buf *const inputs,
+                             const int num_super_frames,
+                             const int remained_spatial_layers) {
     ASSERT_TRUE(inputs != NULL);
     ASSERT_GT(num_super_frames, 0);
     ASSERT_GT(remained_spatial_layers, 0);
@@ -250,45 +249,6 @@ class SvcTest : public ::testing::Test {
       if (frame_count == 0) {
         // There's no super frame but only a single frame.
         ASSERT_EQ(1, remained_spatial_layers);
-        if (is_multiple_frame_contexts) {
-          // Make a new super frame.
-          uint8_t marker = 0xc1;
-          unsigned int mask;
-          int mag;
-
-          // Choose the magnitude.
-          for (mag = 0, mask = 0xff; mag < 4; ++mag) {
-            if (inputs[i].sz < mask)
-              break;
-            mask <<= 8;
-            mask |= 0xff;
-          }
-          marker |= mag << 3;
-          int index_sz = 2 + (mag + 1) * 2;
-
-          inputs[i].buf = realloc(inputs[i].buf, inputs[i].sz + index_sz + 16);
-          ASSERT_TRUE(inputs[i].buf != NULL);
-          uint8_t *frame_data = static_cast<uint8_t*>(inputs[i].buf);
-          frame_data[0] &= ~2;      // Set the show_frame flag to 0.
-          frame_data += inputs[i].sz;
-          // Add an one byte frame with show_existing_frame.
-          *frame_data++ = 0x88;
-
-          // Write the super frame index.
-          *frame_data++ = marker;
-
-          frame_sizes[0] = inputs[i].sz;
-          frame_sizes[1] = 1;
-          for (int j = 0; j < 2; ++j) {
-            unsigned int this_sz = frame_sizes[j];
-            for (int k = 0; k <= mag; k++) {
-              *frame_data++ = this_sz & 0xff;
-              this_sz >>= 8;
-            }
-          }
-          *frame_data++ = marker;
-          inputs[i].sz += index_sz + 1;
-        }
       } else {
         // Found a super frame.
         uint8_t *frame_data = static_cast<uint8_t*>(inputs[i].buf);
@@ -304,16 +264,13 @@ class SvcTest : public ::testing::Test {
         }
         ASSERT_LT(frame, frame_count) << "Couldn't find a visible frame. "
             << "remained_spatial_layers: " << remained_spatial_layers
-            << "    super_frame: " << i
-            << "    is_multiple_frame_context: " << is_multiple_frame_contexts;
-        if (frame == frame_count - 1 && !is_multiple_frame_contexts)
+            << "    super_frame: " << i;
+        if (frame == frame_count - 1)
           continue;
 
         frame_data += frame_sizes[frame];
 
         // We need to add one more frame for multiple frame contexts.
-        if (is_multiple_frame_contexts)
-          ++frame;
         uint8_t marker =
             static_cast<const uint8_t*>(inputs[i].buf)[inputs[i].sz - 1];
         const uint32_t mag = ((marker >> 3) & 0x3) + 1;
@@ -323,35 +280,14 @@ class SvcTest : public ::testing::Test {
         marker |= frame;
 
         // Copy existing frame sizes.
-        memmove(frame_data + (is_multiple_frame_contexts ? 2 : 1),
-                frame_start + inputs[i].sz - index_sz + 1, new_index_sz - 2);
-        if (is_multiple_frame_contexts) {
-          // Add a one byte frame with flag show_existing_frame.
-          *frame_data++ = 0x88 | (remained_spatial_layers - 1);
-        }
+        memmove(frame_data + 1, frame_start + inputs[i].sz - index_sz + 1,
+                new_index_sz - 2);
         // New marker.
         frame_data[0] = marker;
         frame_data += (mag * (frame + 1) + 1);
 
-        if (is_multiple_frame_contexts) {
-          // Write the frame size for the one byte frame.
-          frame_data -= mag;
-          *frame_data++ = 1;
-          for (uint32_t j = 1; j < mag; ++j) {
-            *frame_data++ = 0;
-          }
-        }
-
         *frame_data++ = marker;
         inputs[i].sz = frame_data - frame_start;
-
-        if (is_multiple_frame_contexts) {
-          // Change the show frame flag to 0 for all frames.
-          for (int j = 0; j < frame; ++j) {
-            frame_start[0] &= ~2;
-            frame_start += frame_sizes[j];
-          }
-        }
       }
     }
   }
@@ -555,7 +491,7 @@ TEST_F(SvcTest, TwoPassEncode2SpatialLayersDecodeBaseLayerOnly) {
   vpx_fixed_buf outputs[10];
   memset(&outputs[0], 0, sizeof(outputs));
   Pass2EncodeNFrames(&stats_buf, 10, 2, &outputs[0]);
-  DropLayersAndMakeItVP9Comaptible(&outputs[0], 10, 1, false);
+  DropEnhancementLayers(&outputs[0], 10, 1);
   DecodeNFrames(&outputs[0], 10);
   FreeBitstreamBuffers(&outputs[0], 10);
 }
@@ -573,13 +509,13 @@ TEST_F(SvcTest, TwoPassEncode5SpatialLayersDecode54321Layers) {
   Pass2EncodeNFrames(&stats_buf, 10, 5, &outputs[0]);
 
   DecodeNFrames(&outputs[0], 10);
-  DropLayersAndMakeItVP9Comaptible(&outputs[0], 10, 4, false);
+  DropEnhancementLayers(&outputs[0], 10, 4);
   DecodeNFrames(&outputs[0], 10);
-  DropLayersAndMakeItVP9Comaptible(&outputs[0], 10, 3, false);
+  DropEnhancementLayers(&outputs[0], 10, 3);
   DecodeNFrames(&outputs[0], 10);
-  DropLayersAndMakeItVP9Comaptible(&outputs[0], 10, 2, false);
+  DropEnhancementLayers(&outputs[0], 10, 2);
   DecodeNFrames(&outputs[0], 10);
-  DropLayersAndMakeItVP9Comaptible(&outputs[0], 10, 1, false);
+  DropEnhancementLayers(&outputs[0], 10, 1);
   DecodeNFrames(&outputs[0], 10);
 
   FreeBitstreamBuffers(&outputs[0], 10);
@@ -616,9 +552,9 @@ TEST_F(SvcTest, TwoPassEncode3SNRLayersDecode321Layers) {
   memset(&outputs[0], 0, sizeof(outputs));
   Pass2EncodeNFrames(&stats_buf, 20, 3, &outputs[0]);
   DecodeNFrames(&outputs[0], 20);
-  DropLayersAndMakeItVP9Comaptible(&outputs[0], 20, 2, false);
+  DropEnhancementLayers(&outputs[0], 20, 2);
   DecodeNFrames(&outputs[0], 20);
-  DropLayersAndMakeItVP9Comaptible(&outputs[0], 20, 1, false);
+  DropEnhancementLayers(&outputs[0], 20, 1);
   DecodeNFrames(&outputs[0], 20);
 
   FreeBitstreamBuffers(&outputs[0], 20);
@@ -649,7 +585,6 @@ TEST_F(SvcTest, TwoPassEncode2SpatialLayersWithMultipleFrameContexts) {
   vpx_fixed_buf outputs[10];
   memset(&outputs[0], 0, sizeof(outputs));
   Pass2EncodeNFrames(&stats_buf, 10, 2, &outputs[0]);
-  DropLayersAndMakeItVP9Comaptible(&outputs[0], 10, 2, true);
   DecodeNFrames(&outputs[0], 10);
   FreeBitstreamBuffers(&outputs[0], 10);
 }
@@ -667,7 +602,7 @@ TEST_F(SvcTest,
   vpx_fixed_buf outputs[10];
   memset(&outputs[0], 0, sizeof(outputs));
   Pass2EncodeNFrames(&stats_buf, 10, 2, &outputs[0]);
-  DropLayersAndMakeItVP9Comaptible(&outputs[0], 10, 1, true);
+  DropEnhancementLayers(&outputs[0], 10, 1);
   DecodeNFrames(&outputs[0], 10);
   FreeBitstreamBuffers(&outputs[0], 10);
 }
@@ -686,7 +621,6 @@ TEST_F(SvcTest, TwoPassEncode2SNRLayersWithMultipleFrameContexts) {
   vpx_fixed_buf outputs[10];
   memset(&outputs[0], 0, sizeof(outputs));
   Pass2EncodeNFrames(&stats_buf, 10, 2, &outputs[0]);
-  DropLayersAndMakeItVP9Comaptible(&outputs[0], 10, 2, true);
   DecodeNFrames(&outputs[0], 10);
   FreeBitstreamBuffers(&outputs[0], 10);
 }
@@ -707,32 +641,13 @@ TEST_F(SvcTest,
   memset(&outputs[0], 0, sizeof(outputs));
   Pass2EncodeNFrames(&stats_buf, 10, 3, &outputs[0]);
 
-  vpx_fixed_buf outputs_new[10];
-  for (int i = 0; i < 10; ++i) {
-    outputs_new[i].buf = malloc(outputs[i].sz + 16);
-    ASSERT_TRUE(outputs_new[i].buf != NULL);
-    memcpy(outputs_new[i].buf, outputs[i].buf, outputs[i].sz);
-    outputs_new[i].sz = outputs[i].sz;
-  }
-  DropLayersAndMakeItVP9Comaptible(&outputs_new[0], 10, 3, true);
-  DecodeNFrames(&outputs_new[0], 10);
-
-  for (int i = 0; i < 10; ++i) {
-    memcpy(outputs_new[i].buf, outputs[i].buf, outputs[i].sz);
-    outputs_new[i].sz = outputs[i].sz;
-  }
-  DropLayersAndMakeItVP9Comaptible(&outputs_new[0], 10, 2, true);
-  DecodeNFrames(&outputs_new[0], 10);
-
-  for (int i = 0; i < 10; ++i) {
-    memcpy(outputs_new[i].buf, outputs[i].buf, outputs[i].sz);
-    outputs_new[i].sz = outputs[i].sz;
-  }
-  DropLayersAndMakeItVP9Comaptible(&outputs_new[0], 10, 1, true);
-  DecodeNFrames(&outputs_new[0], 10);
+  DecodeNFrames(&outputs[0], 10);
+  DropEnhancementLayers(&outputs[0], 10, 2);
+  DecodeNFrames(&outputs[0], 10);
+  DropEnhancementLayers(&outputs[0], 10, 1);
+  DecodeNFrames(&outputs[0], 10);
 
   FreeBitstreamBuffers(&outputs[0], 10);
-  FreeBitstreamBuffers(&outputs_new[0], 10);
 }
 
 TEST_F(SvcTest, TwoPassEncode2TemporalLayers) {
@@ -769,7 +684,6 @@ TEST_F(SvcTest, TwoPassEncode2TemporalLayersWithMultipleFrameContexts) {
   vpx_fixed_buf outputs[10];
   memset(&outputs[0], 0, sizeof(outputs));
   Pass2EncodeNFrames(&stats_buf, 10, 1, &outputs[0]);
-  DropLayersAndMakeItVP9Comaptible(&outputs[0], 10, 1, true);
   DecodeNFrames(&outputs[0], 10);
   FreeBitstreamBuffers(&outputs[0], 10);
 }
@@ -814,7 +728,6 @@ TEST_F(SvcTest,
   vpx_fixed_buf outputs[10];
   memset(&outputs[0], 0, sizeof(outputs));
   Pass2EncodeNFrames(&stats_buf, 10, 1, &outputs[0]);
-  DropLayersAndMakeItVP9Comaptible(&outputs[0], 10, 1, true);
 
   vpx_fixed_buf base_layer[5];
   for (int i = 0; i < 5; ++i)
diff --git a/vp9/encoder/vp9_bitstream.c b/vp9/encoder/vp9_bitstream.c
index 694cac76f..3954fe6a7 100644
--- a/vp9/encoder/vp9_bitstream.c
+++ b/vp9/encoder/vp9_bitstream.c
@@ -1013,7 +1013,11 @@ static void write_frame_size_with_refs(VP9_COMP *cpi,
         ((cpi->svc.number_temporal_layers > 1 &&
          cpi->oxcf.rc_mode == VPX_CBR) ||
         (cpi->svc.number_spatial_layers > 1 &&
-         cpi->svc.layer_context[cpi->svc.spatial_layer_id].is_key_frame))) {
+         cpi->svc.layer_context[cpi->svc.spatial_layer_id].is_key_frame) ||
+        (is_two_pass_svc(cpi) &&
+         cpi->svc.encode_empty_frame_state == ENCODING &&
+         cpi->svc.layer_context[0].frames_from_key_frame <
+         cpi->svc.number_temporal_layers + 1))) {
       found = 0;
     }
     vp9_wb_write_bit(wb, found);
@@ -1105,8 +1109,7 @@ static void write_uncompressed_header(VP9_COMP *cpi,
     // will change to show_frame flag to 0, then add an one byte frame with
     // show_existing_frame flag which tells the decoder which frame we want to
     // show.
-    if (!cm->show_frame ||
-        (is_two_pass_svc(cpi) && cm->error_resilient_mode == 0))
+    if (!cm->show_frame)
       vp9_wb_write_bit(wb, cm->intra_only);
 
     if (!cm->error_resilient_mode)
diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index 430cd0004..1758e3fdb 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -225,6 +225,9 @@ static void dealloc_compressor_data(VP9_COMP *cpi) {
   }
   vpx_memset(&cpi->svc.scaled_frames[0], 0,
              MAX_LAG_BUFFERS * sizeof(cpi->svc.scaled_frames[0]));
+
+  vp9_free_frame_buffer(&cpi->svc.empty_frame.img);
+  vpx_memset(&cpi->svc.empty_frame, 0, sizeof(cpi->svc.empty_frame));
 }
 
 static void save_coding_context(VP9_COMP *cpi) {
@@ -2979,7 +2982,9 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
     }
   }
   if (is_two_pass_svc(cpi) && cm->error_resilient_mode == 0) {
+    // Use the last frame context for the empty frame.
     cm->frame_context_idx =
+        (cpi->svc.encode_empty_frame_state == ENCODING) ? FRAME_CONTEXTS - 1 :
         cpi->svc.spatial_layer_id * cpi->svc.number_temporal_layers +
         cpi->svc.temporal_layer_id;
 
@@ -3160,7 +3165,9 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
   cpi->ref_frame_flags = get_ref_frame_flags(cpi);
 
   cm->last_frame_type = cm->frame_type;
-  vp9_rc_postencode_update(cpi, *size);
+
+  if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING))
+    vp9_rc_postencode_update(cpi, *size);
 
 #if 0
   output_frame_level_debug_stats(cpi);
@@ -3184,12 +3191,8 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
   cm->last_height = cm->height;
 
   // reset to normal state now that we are done.
-  if (!cm->show_existing_frame) {
-    if (is_two_pass_svc(cpi) && cm->error_resilient_mode == 0)
-      cm->last_show_frame = 0;
-    else
-      cm->last_show_frame = cm->show_frame;
-  }
+  if (!cm->show_existing_frame)
+    cm->last_show_frame = cm->show_frame;
 
   if (cm->show_frame) {
     vp9_swap_mi_and_prev_mi(cm);
@@ -3226,7 +3229,9 @@ static void Pass2Encode(VP9_COMP *cpi, size_t *size,
                         uint8_t *dest, unsigned int *frame_flags) {
   cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
   encode_frame_to_data_rate(cpi, size, dest, frame_flags);
-  vp9_twopass_postencode_update(cpi);
+
+  if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING))
+    vp9_twopass_postencode_update(cpi);
 }
 
 static void init_motion_estimation(VP9_COMP *cpi) {
@@ -3414,6 +3419,9 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
   if (is_two_pass_svc(cpi)) {
 #if CONFIG_SPATIAL_SVC
     vp9_svc_start_frame(cpi);
+    // Use a small empty frame instead of a real frame
+    if (cpi->svc.encode_empty_frame_state == ENCODING)
+      source = &cpi->svc.empty_frame;
 #endif
     if (oxcf->pass == 2)
       vp9_restore_layer_context(cpi);
@@ -3432,6 +3440,11 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
 
   // Should we encode an arf frame.
   arf_src_index = get_arf_src_index(cpi);
+
+  // Skip alt frame if we encode the empty frame
+  if (is_two_pass_svc(cpi) && source != NULL)
+    arf_src_index = 0;
+
   if (arf_src_index) {
     assert(arf_src_index <= rc->frames_to_key);
 
@@ -3542,7 +3555,10 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
 
   // For two pass encodes analyse the first pass stats and determine
   // the bit allocation and other parameters for this frame / group of frames.
-  if ((oxcf->pass == 2) && (!cpi->use_svc || is_two_pass_svc(cpi))) {
+  if ((oxcf->pass == 2) &&
+      (!cpi->use_svc ||
+       (is_two_pass_svc(cpi) &&
+        cpi->svc.encode_empty_frame_state != ENCODING))) {
     vp9_rc_get_second_pass_params(cpi);
   }
 
@@ -3771,10 +3787,18 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
 
 #endif
 
-  if (is_two_pass_svc(cpi) && cm->show_frame) {
-    ++cpi->svc.spatial_layer_to_encode;
-    if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
-      cpi->svc.spatial_layer_to_encode = 0;
+  if (is_two_pass_svc(cpi)) {
+    if (cpi->svc.encode_empty_frame_state == ENCODING)
+      cpi->svc.encode_empty_frame_state = ENCODED;
+
+    if (cm->show_frame) {
+      ++cpi->svc.spatial_layer_to_encode;
+      if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
+        cpi->svc.spatial_layer_to_encode = 0;
+
+      // May need the empty frame after an visible frame.
+      cpi->svc.encode_empty_frame_state = NEED_TO_ENCODE;
+    }
   }
   return 0;
 }
@@ -3865,10 +3889,6 @@ int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
 
   if (width) {
     cm->width = width;
-    if (cm->width * 5 < cpi->initial_width) {
-      cm->width = cpi->initial_width / 5 + 1;
-      printf("Warning: Desired width too small, changed to %d\n", cm->width);
-    }
     if (cm->width > cpi->initial_width) {
       cm->width = cpi->initial_width;
       printf("Warning: Desired width too large, changed to %d\n", cm->width);
@@ -3877,10 +3897,6 @@ int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
 
   if (height) {
     cm->height = height;
-    if (cm->height * 5 < cpi->initial_height) {
-      cm->height = cpi->initial_height / 5 + 1;
-      printf("Warning: Desired height too small, changed to %d\n", cm->height);
-    }
     if (cm->height > cpi->initial_height) {
       cm->height = cpi->initial_height;
       printf("Warning: Desired height too large, changed to %d\n", cm->height);
diff --git a/vp9/encoder/vp9_firstpass.c b/vp9/encoder/vp9_firstpass.c
index 98daa17fa..96c3e0aa4 100644
--- a/vp9/encoder/vp9_firstpass.c
+++ b/vp9/encoder/vp9_firstpass.c
@@ -2405,6 +2405,9 @@ void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
         cpi->ref_frame_flags &=
             (~VP9_LAST_FLAG & ~VP9_GOLD_FLAG & ~VP9_ALT_FLAG);
         lc->frames_from_key_frame = 0;
+        // Reset the empty frame resolution since we have a key frame.
+        cpi->svc.empty_frame_width = cm->width;
+        cpi->svc.empty_frame_height = cm->height;
       }
     } else {
       cm->frame_type = INTER_FRAME;
diff --git a/vp9/encoder/vp9_svc_layercontext.c b/vp9/encoder/vp9_svc_layercontext.c
index eed681c96..1573557d4 100644
--- a/vp9/encoder/vp9_svc_layercontext.c
+++ b/vp9/encoder/vp9_svc_layercontext.c
@@ -14,6 +14,8 @@
 #include "vp9/encoder/vp9_svc_layercontext.h"
 #include "vp9/encoder/vp9_extend.h"
 
+#define SMALL_FRAME_FB_IDX 7
+
 void vp9_init_layer_context(VP9_COMP *const cpi) {
   SVC *const svc = &cpi->svc;
   const VP9EncoderConfig *const oxcf = &cpi->oxcf;
@@ -28,6 +30,25 @@ void vp9_init_layer_context(VP9_COMP *const cpi) {
     layer_end = svc->number_temporal_layers;
   } else {
     layer_end = svc->number_spatial_layers;
+
+    if (cpi->oxcf.error_resilient_mode == 0 && cpi->oxcf.pass == 2) {
+      if (vp9_realloc_frame_buffer(&cpi->svc.empty_frame.img,
+                                   cpi->common.width, cpi->common.height,
+                                   cpi->common.subsampling_x,
+                                   cpi->common.subsampling_y,
+#if CONFIG_VP9_HIGHBITDEPTH
+                                 cpi->common.use_highbitdepth,
+#endif
+                                 VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL))
+        vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
+                           "Failed to allocate empty frame for multiple frame "
+                           "contexts");
+
+      vpx_memset(cpi->svc.empty_frame.img.buffer_alloc, 0x80,
+                 cpi->svc.empty_frame.img.buffer_alloc_sz);
+      cpi->svc.empty_frame_width = cpi->common.width;
+      cpi->svc.empty_frame_height = cpi->common.height;
+    }
   }
 
   for (layer = 0; layer < layer_end; ++layer) {
@@ -310,6 +331,47 @@ int vp9_svc_start_frame(VP9_COMP *const cpi) {
   get_layer_resolution(cpi->oxcf.width, cpi->oxcf.height,
                        lc->scaling_factor_num, lc->scaling_factor_den,
                        &width, &height);
+
+  // Workaround for multiple frame contexts. In some frames we can't use prev_mi
+  // since its previous frame could be changed during decoding time. The idea is
+  // we put a empty invisible frame in front of them, then we will not use
+  // prev_mi when encoding these frames.
+  if (cpi->oxcf.error_resilient_mode == 0 && cpi->oxcf.pass == 2 &&
+      cpi->svc.encode_empty_frame_state == NEED_TO_ENCODE) {
+    if ((cpi->svc.number_temporal_layers > 1 &&
+         cpi->svc.temporal_layer_id < cpi->svc.number_temporal_layers - 1) ||
+        (cpi->svc.number_spatial_layers > 1 &&
+         cpi->svc.spatial_layer_id == 0)) {
+      struct lookahead_entry *buf = vp9_lookahead_peek(cpi->lookahead, 0);
+
+      if (buf != NULL) {
+        cpi->svc.empty_frame.ts_start = buf->ts_start;
+        cpi->svc.empty_frame.ts_end = buf->ts_end;
+        cpi->svc.encode_empty_frame_state = ENCODING;
+        cpi->common.show_frame = 0;
+        cpi->ref_frame_flags = 0;
+        cpi->common.frame_type = INTER_FRAME;
+        cpi->lst_fb_idx =
+            cpi->gld_fb_idx = cpi->alt_fb_idx = SMALL_FRAME_FB_IDX;
+
+        // Gradually make the empty frame smaller to save bits. Make it half of
+        // its previous size because of the scaling factor restriction.
+        cpi->svc.empty_frame_width >>= 1;
+        cpi->svc.empty_frame_width = (cpi->svc.empty_frame_width + 1) & ~1;
+        if (cpi->svc.empty_frame_width < 16)
+          cpi->svc.empty_frame_width = 16;
+
+        cpi->svc.empty_frame_height >>= 1;
+        cpi->svc.empty_frame_height = (cpi->svc.empty_frame_height + 1) & ~1;
+        if (cpi->svc.empty_frame_height < 16)
+          cpi->svc.empty_frame_height = 16;
+
+        width = cpi->svc.empty_frame_width;
+        height = cpi->svc.empty_frame_height;
+      }
+    }
+  }
+
   if (vp9_set_size_literal(cpi, width, height) != 0)
     return VPX_CODEC_INVALID_PARAM;
 
@@ -317,7 +379,6 @@ int vp9_svc_start_frame(VP9_COMP *const cpi) {
   cpi->oxcf.best_allowed_q = vp9_quantizer_to_qindex(lc->min_q);
 
   vp9_change_config(cpi, &cpi->oxcf);
-
   vp9_set_high_precision_mv(cpi, 1);
 
   cpi->alt_ref_source = get_layer_context(cpi)->alt_ref_source;
diff --git a/vp9/encoder/vp9_svc_layercontext.h b/vp9/encoder/vp9_svc_layercontext.h
index 47a5456b6..e9645ce9f 100644
--- a/vp9/encoder/vp9_svc_layercontext.h
+++ b/vp9/encoder/vp9_svc_layercontext.h
@@ -50,6 +50,16 @@ typedef struct {
 
   int spatial_layer_to_encode;
 
+  // Workaround for multiple frame contexts
+  enum {
+    ENCODED = 0,
+    ENCODING,
+    NEED_TO_ENCODE
+  }encode_empty_frame_state;
+  struct lookahead_entry empty_frame;
+  int empty_frame_width;
+  int empty_frame_height;
+
   // Store scaled source frames to be used for temporal filter to generate
   // a alt ref frame.
   YV12_BUFFER_CONFIG scaled_frames[MAX_LAG_BUFFERS];
diff --git a/vp9/encoder/vp9_temporal_filter.c b/vp9/encoder/vp9_temporal_filter.c
index 9ae81e761..5599227ce 100644
--- a/vp9/encoder/vp9_temporal_filter.c
+++ b/vp9/encoder/vp9_temporal_filter.c
@@ -719,6 +719,9 @@ void vp9_temporal_filter(VP9_COMP *cpi, int distance) {
           ++frame_used;
         }
       }
+      cm->mi = cm->mip + cm->mi_stride + 1;
+      cpi->mb.e_mbd.mi = cm->mi;
+      cpi->mb.e_mbd.mi[0].src_mi = &cpi->mb.e_mbd.mi[0];
     } else {
       // ARF is produced at the native frame size and resized when coded.
 #if CONFIG_VP9_HIGHBITDEPTH
diff --git a/vp9/vp9_cx_iface.c b/vp9/vp9_cx_iface.c
index 041ba27da..d0ca5242c 100644
--- a/vp9/vp9_cx_iface.c
+++ b/vp9/vp9_cx_iface.c
@@ -188,11 +188,9 @@ static vpx_codec_err_t validate_config(vpx_codec_alg_priv_t *ctx,
     }
     if (alt_ref_sum > REF_FRAMES - cfg->ss_number_layers)
       ERROR("Not enough ref buffers for svc alt ref frames");
-    if ((cfg->ss_number_layers > 3 ||
-         cfg->ss_number_layers * cfg->ts_number_layers > 4) &&
+    if (cfg->ss_number_layers * cfg->ts_number_layers > 3 &&
         cfg->g_error_resilient == 0)
-    ERROR("Multiple frame context are not supported for more than 3 spatial "
-          "layers or more than 4 spatial x temporal layers");
+    ERROR("Multiple frame context are not supported for more than 3 layers");
   }
 #endif
 
diff --git a/vpx/src/svc_encodeframe.c b/vpx/src/svc_encodeframe.c
index 773087ddc..fa3409c69 100644
--- a/vpx/src/svc_encodeframe.c
+++ b/vpx/src/svc_encodeframe.c
@@ -350,7 +350,7 @@ void assign_layer_bitrates(const SvcContext *svc_ctx,
       }
     }
 
-    for (i = 0; i < svc_ctx->spatial_layers; ++i) {
+    for (i = 0; i < VPX_SS_MAX_LAYERS; ++i) {
       if (total > 0) {
         enc_cfg->ss_target_bitrate[i] = (unsigned int)
             (enc_cfg->rc_target_bitrate * alloc_ratio[i] / total);