diff --git a/test/byte_alignment_test.cc b/test/byte_alignment_test.cc
new file mode 100644
index 000000000..aa4b78b9a
--- /dev/null
+++ b/test/byte_alignment_test.cc
@@ -0,0 +1,189 @@
+/*
+ *  Copyright (c) 2014 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 <string>
+
+#include "./vpx_config.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#if CONFIG_WEBM_IO
+#include "test/webm_video_source.h"
+#endif
+
+namespace {
+
+const int kLegacyByteAlignment = 0;
+const int kLegacyYPlaneByteAlignment = 32;
+const int kNumPlanesToCheck = 3;
+const char kVP9TestFile[] = "vp90-2-02-size-lf-1920x1080.webm";
+const char kVP9Md5File[] = "vp90-2-02-size-lf-1920x1080.webm.md5";
+
+#if CONFIG_WEBM_IO
+
+struct ByteAlignmentTestParam {
+  int byte_alignment;
+  vpx_codec_err_t expected_value;
+  bool decode_remaining;
+};
+
+const ByteAlignmentTestParam kBaTestParams[] = {
+  {kLegacyByteAlignment, VPX_CODEC_OK, true},
+  {32, VPX_CODEC_OK, true},
+  {64, VPX_CODEC_OK, true},
+  {128, VPX_CODEC_OK, true},
+  {256, VPX_CODEC_OK, true},
+  {512, VPX_CODEC_OK, true},
+  {1024, VPX_CODEC_OK, true},
+  {1, VPX_CODEC_INVALID_PARAM, false},
+  {-2, VPX_CODEC_INVALID_PARAM, false},
+  {4, VPX_CODEC_INVALID_PARAM, false},
+  {16, VPX_CODEC_INVALID_PARAM, false},
+  {255, VPX_CODEC_INVALID_PARAM, false},
+  {2048, VPX_CODEC_INVALID_PARAM, false},
+};
+
+// Class for testing byte alignment of reference buffers.
+class ByteAlignmentTest
+    : public ::testing::TestWithParam<ByteAlignmentTestParam> {
+ protected:
+  ByteAlignmentTest()
+      : video_(NULL),
+        decoder_(NULL),
+        md5_file_(NULL) {}
+
+  virtual void SetUp() {
+    video_ = new libvpx_test::WebMVideoSource(kVP9TestFile);
+    ASSERT_TRUE(video_ != NULL);
+    video_->Init();
+    video_->Begin();
+
+    const vpx_codec_dec_cfg_t cfg = vpx_codec_dec_cfg_t();
+    decoder_ = new libvpx_test::VP9Decoder(cfg, 0);
+    ASSERT_TRUE(decoder_ != NULL);
+
+    OpenMd5File(kVP9Md5File);
+  }
+
+  virtual void TearDown() {
+    if (md5_file_ != NULL)
+      fclose(md5_file_);
+
+    delete decoder_;
+    delete video_;
+  }
+
+  void SetByteAlignment(int byte_alignment, vpx_codec_err_t expected_value) {
+    decoder_->Control(VP9_SET_BYTE_ALIGNMENT, byte_alignment, expected_value);
+  }
+
+  vpx_codec_err_t DecodeOneFrame(int byte_alignment_to_check) {
+    const vpx_codec_err_t res =
+        decoder_->DecodeFrame(video_->cxdata(), video_->frame_size());
+    CheckDecodedFrames(byte_alignment_to_check);
+    if (res == VPX_CODEC_OK)
+      video_->Next();
+    return res;
+  }
+
+  vpx_codec_err_t DecodeRemainingFrames(int byte_alignment_to_check) {
+    for (; video_->cxdata() != NULL; video_->Next()) {
+      const vpx_codec_err_t res =
+          decoder_->DecodeFrame(video_->cxdata(), video_->frame_size());
+      if (res != VPX_CODEC_OK)
+        return res;
+      CheckDecodedFrames(byte_alignment_to_check);
+    }
+    return VPX_CODEC_OK;
+  }
+
+ private:
+  // Check if |data| is aligned to |byte_alignment_to_check|.
+  // |byte_alignment_to_check| must be a power of 2.
+  void CheckByteAlignment(const uint8_t *data, int byte_alignment_to_check) {
+    ASSERT_EQ(0u, reinterpret_cast<size_t>(data) % byte_alignment_to_check);
+  }
+
+  // Iterate through the planes of the decoded frames and check for
+  // alignment based off |byte_alignment_to_check|.
+  void CheckDecodedFrames(int byte_alignment_to_check) {
+    libvpx_test::DxDataIterator dec_iter = decoder_->GetDxData();
+    const vpx_image_t *img;
+
+    // Get decompressed data
+    while ((img = dec_iter.Next()) != NULL) {
+      if (byte_alignment_to_check == kLegacyByteAlignment) {
+        CheckByteAlignment(img->planes[0], kLegacyYPlaneByteAlignment);
+      } else {
+        for (int i = 0; i < kNumPlanesToCheck; ++i) {
+          CheckByteAlignment(img->planes[i], byte_alignment_to_check);
+        }
+      }
+      CheckMd5(*img);
+    }
+  }
+
+  // TODO(fgalligan): Move the MD5 testing code into another class.
+  void OpenMd5File(const std::string &md5_file_name_) {
+    md5_file_ = libvpx_test::OpenTestDataFile(md5_file_name_);
+    ASSERT_TRUE(md5_file_ != NULL) << "MD5 file open failed. Filename: "
+        << md5_file_name_;
+  }
+
+  void CheckMd5(const vpx_image_t &img) {
+    ASSERT_TRUE(md5_file_ != NULL);
+    char expected_md5[33];
+    char junk[128];
+
+    // Read correct md5 checksums.
+    const int res = fscanf(md5_file_, "%s  %s", expected_md5, junk);
+    ASSERT_NE(EOF, res) << "Read md5 data failed";
+    expected_md5[32] = '\0';
+
+    ::libvpx_test::MD5 md5_res;
+    md5_res.Add(&img);
+    const char *const actual_md5 = md5_res.Get();
+
+    // Check md5 match.
+    ASSERT_STREQ(expected_md5, actual_md5) << "MD5 checksums don't match";
+  }
+
+  libvpx_test::WebMVideoSource *video_;
+  libvpx_test::VP9Decoder *decoder_;
+  FILE *md5_file_;
+};
+
+TEST_F(ByteAlignmentTest, SwitchByteAlignment) {
+  const int num_elements = 14;
+  const int byte_alignments[] = { 0, 32, 64, 128, 256, 512, 1024,
+                                  0, 1024, 32, 512, 64, 256, 128 };
+
+  for (int i = 0; i < num_elements; ++i) {
+    SetByteAlignment(byte_alignments[i], VPX_CODEC_OK);
+    ASSERT_EQ(VPX_CODEC_OK, DecodeOneFrame(byte_alignments[i]));
+  }
+  SetByteAlignment(byte_alignments[0], VPX_CODEC_OK);
+  ASSERT_EQ(VPX_CODEC_OK, DecodeRemainingFrames(byte_alignments[0]));
+}
+
+TEST_P(ByteAlignmentTest, TestAlignment) {
+  const ByteAlignmentTestParam t = GetParam();
+  SetByteAlignment(t.byte_alignment, t.expected_value);
+  if (t.decode_remaining)
+    ASSERT_EQ(VPX_CODEC_OK, DecodeRemainingFrames(t.byte_alignment));
+}
+
+INSTANTIATE_TEST_CASE_P(Alignments, ByteAlignmentTest,
+                        ::testing::ValuesIn(kBaTestParams));
+
+#endif  // CONFIG_WEBM_IO
+
+}  // namespace
diff --git a/test/decode_test_driver.h b/test/decode_test_driver.h
index 76cd12e02..72a279f00 100644
--- a/test/decode_test_driver.h
+++ b/test/decode_test_driver.h
@@ -66,9 +66,7 @@ class Decoder {
   }
 
   void Control(int ctrl_id, int arg) {
-    InitOnce();
-    const vpx_codec_err_t res = vpx_codec_control_(&decoder_, ctrl_id, arg);
-    ASSERT_EQ(VPX_CODEC_OK, res) << DecodeError();
+    Control(ctrl_id, arg, VPX_CODEC_OK);
   }
 
   void Control(int ctrl_id, const void *arg) {
@@ -77,6 +75,12 @@ class Decoder {
     ASSERT_EQ(VPX_CODEC_OK, res) << DecodeError();
   }
 
+  void Control(int ctrl_id, int arg, vpx_codec_err_t expected_value) {
+    InitOnce();
+    const vpx_codec_err_t res = vpx_codec_control_(&decoder_, ctrl_id, arg);
+    ASSERT_EQ(expected_value, res) << DecodeError();
+  }
+
   const char* DecodeError() {
     const char *detail = vpx_codec_error_detail(&decoder_);
     return detail ? detail : vpx_codec_error(&decoder_);
diff --git a/test/test.mk b/test/test.mk
index d94329924..f5c9c37a7 100644
--- a/test/test.mk
+++ b/test/test.mk
@@ -31,6 +31,7 @@ LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += config_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += cq_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += keyframe_test.cc
 
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += byte_alignment_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += external_frame_buffer_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += invalid_file_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += user_priv_test.cc
diff --git a/vp9/common/vp9_alloccommon.c b/vp9/common/vp9_alloccommon.c
index cb299f9f7..e87f42419 100644
--- a/vp9/common/vp9_alloccommon.c
+++ b/vp9/common/vp9_alloccommon.c
@@ -110,7 +110,8 @@ int vp9_alloc_ref_frame_buffers(VP9_COMMON *cm, int width, int height) {
 #if CONFIG_VP9_HIGHBITDEPTH
                                cm->use_highbitdepth,
 #endif
-                               VP9_ENC_BORDER_IN_PIXELS) < 0)
+                               VP9_ENC_BORDER_IN_PIXELS,
+                               cm->byte_alignment) < 0)
       goto fail;
     if (cm->frame_bufs[i].mvs == NULL) {
       cm->frame_bufs[i].mvs =
diff --git a/vp9/common/vp9_onyxc_int.h b/vp9/common/vp9_onyxc_int.h
index 55a1f86c7..a7e7ab16d 100644
--- a/vp9/common/vp9_onyxc_int.h
+++ b/vp9/common/vp9_onyxc_int.h
@@ -207,6 +207,7 @@ typedef struct VP9Common {
   int frame_parallel_decoding_mode;
 
   int log2_tile_cols, log2_tile_rows;
+  int byte_alignment;
 
   // Private data associated with the frame buffer callbacks.
   void *cb_priv;
diff --git a/vp9/decoder/vp9_decodeframe.c b/vp9/decoder/vp9_decodeframe.c
index 58df87d0c..6eb9026a9 100644
--- a/vp9/decoder/vp9_decodeframe.c
+++ b/vp9/decoder/vp9_decodeframe.c
@@ -719,6 +719,7 @@ static void setup_frame_size(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
           cm->use_highbitdepth,
 #endif
           VP9_DEC_BORDER_IN_PIXELS,
+          cm->byte_alignment,
           &cm->frame_bufs[cm->new_fb_idx].raw_frame_buffer, cm->get_fb_cb,
           cm->cb_priv)) {
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
@@ -793,6 +794,7 @@ static void setup_frame_size_with_refs(VP9_COMMON *cm,
           cm->use_highbitdepth,
 #endif
           VP9_DEC_BORDER_IN_PIXELS,
+          cm->byte_alignment,
           &cm->frame_bufs[cm->new_fb_idx].raw_frame_buffer, cm->get_fb_cb,
           cm->cb_priv)) {
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
diff --git a/vp9/encoder/vp9_denoiser.c b/vp9/encoder/vp9_denoiser.c
index 4deeed217..56ec6b335 100644
--- a/vp9/encoder/vp9_denoiser.c
+++ b/vp9/encoder/vp9_denoiser.c
@@ -425,6 +425,7 @@ int vp9_denoiser_alloc(VP9_DENOISER *denoiser, int width, int height,
 #endif
                        int border) {
   int i, fail;
+  const int legacy_byte_alignment = 0;
   assert(denoiser != NULL);
 
   for (i = 0; i < MAX_REF_FRAMES; ++i) {
@@ -433,7 +434,7 @@ int vp9_denoiser_alloc(VP9_DENOISER *denoiser, int width, int height,
 #if CONFIG_VP9_HIGHBITDEPTH
                                   use_highbitdepth,
 #endif
-                                  border);
+                                  border, legacy_byte_alignment);
     if (fail) {
       vp9_denoiser_free(denoiser);
       return 1;
@@ -448,7 +449,7 @@ int vp9_denoiser_alloc(VP9_DENOISER *denoiser, int width, int height,
 #if CONFIG_VP9_HIGHBITDEPTH
                                 use_highbitdepth,
 #endif
-                                border);
+                                border, legacy_byte_alignment);
   if (fail) {
     vp9_denoiser_free(denoiser);
     return 1;
diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index aaa6b238d..a7f34a2e4 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -491,7 +491,8 @@ static void alloc_raw_frame_buffers(VP9_COMP *cpi) {
 #if CONFIG_VP9_HIGHBITDEPTH
                                cm->use_highbitdepth,
 #endif
-                               VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL))
+                               VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
+                               NULL, NULL, NULL))
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                        "Failed to allocate altref buffer");
 }
@@ -511,7 +512,8 @@ static void alloc_util_frame_buffers(VP9_COMP *cpi) {
 #if CONFIG_VP9_HIGHBITDEPTH
                                cm->use_highbitdepth,
 #endif
-                               VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL))
+                               VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
+                               NULL, NULL, NULL))
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                        "Failed to allocate last frame buffer");
 
@@ -521,7 +523,8 @@ static void alloc_util_frame_buffers(VP9_COMP *cpi) {
 #if CONFIG_VP9_HIGHBITDEPTH
                                cm->use_highbitdepth,
 #endif
-                               VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL))
+                               VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
+                               NULL, NULL, NULL))
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                        "Failed to allocate scaled source buffer");
 
@@ -531,7 +534,8 @@ static void alloc_util_frame_buffers(VP9_COMP *cpi) {
 #if CONFIG_VP9_HIGHBITDEPTH
                                cm->use_highbitdepth,
 #endif
-                               VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL))
+                               VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
+                               NULL, NULL, NULL))
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                        "Failed to allocate scaled last source buffer");
 }
@@ -567,7 +571,8 @@ static void update_frame_size(VP9_COMP *cpi) {
 #if CONFIG_VP9_HIGHBITDEPTH
                                  cm->use_highbitdepth,
 #endif
-                                 VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL))
+                                 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
+                                 NULL, NULL, NULL))
       vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                          "Failed to reallocate alt_ref_buffer");
   }
@@ -2473,7 +2478,8 @@ void vp9_scale_references(VP9_COMP *cpi) {
                                  cm->width, cm->height,
                                  cm->subsampling_x, cm->subsampling_y,
                                  cm->use_highbitdepth,
-                                 VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL);
+                                 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
+                                 NULL, NULL, NULL);
         scale_and_extend_frame(ref, &cm->frame_bufs[new_fb].buf,
                                (int)cm->bit_depth);
 #else
@@ -2482,7 +2488,8 @@ void vp9_scale_references(VP9_COMP *cpi) {
         vp9_realloc_frame_buffer(&cm->frame_bufs[new_fb].buf,
                                  cm->width, cm->height,
                                  cm->subsampling_x, cm->subsampling_y,
-                                 VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL);
+                                 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
+                                 NULL, NULL, NULL);
         scale_and_extend_frame(ref, &cm->frame_bufs[new_fb].buf);
 #endif  // CONFIG_VP9_HIGHBITDEPTH
         cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
@@ -2721,7 +2728,8 @@ void set_frame_size(VP9_COMP *cpi) {
 #if CONFIG_VP9_HIGHBITDEPTH
                            cm->use_highbitdepth,
 #endif
-                           VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL);
+                           VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
+                           NULL, NULL, NULL);
 
   alloc_util_frame_buffers(cpi);
   init_motion_estimation(cpi);
diff --git a/vp9/encoder/vp9_lookahead.c b/vp9/encoder/vp9_lookahead.c
index 823e7a162..708072ee2 100644
--- a/vp9/encoder/vp9_lookahead.c
+++ b/vp9/encoder/vp9_lookahead.c
@@ -65,6 +65,7 @@ struct lookahead_ctx *vp9_lookahead_init(unsigned int width,
   // Allocate the lookahead structures
   ctx = calloc(1, sizeof(*ctx));
   if (ctx) {
+    const int legacy_byte_alignment = 0;
     unsigned int i;
     ctx->max_sz = depth;
     ctx->buf = calloc(depth, sizeof(*ctx->buf));
@@ -76,7 +77,8 @@ struct lookahead_ctx *vp9_lookahead_init(unsigned int width,
 #if CONFIG_VP9_HIGHBITDEPTH
                                  use_highbitdepth,
 #endif
-                                 VP9_ENC_BORDER_IN_PIXELS))
+                                 VP9_ENC_BORDER_IN_PIXELS,
+                                 legacy_byte_alignment))
         goto bail;
   }
   return ctx;
diff --git a/vp9/encoder/vp9_svc_layercontext.c b/vp9/encoder/vp9_svc_layercontext.c
index 184322f4f..31e93be65 100644
--- a/vp9/encoder/vp9_svc_layercontext.c
+++ b/vp9/encoder/vp9_svc_layercontext.c
@@ -39,7 +39,9 @@ void vp9_init_layer_context(VP9_COMP *const cpi) {
 #if CONFIG_VP9_HIGHBITDEPTH
                                  cpi->common.use_highbitdepth,
 #endif
-                                 VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL))
+                                 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 empty frame for multiple frame "
                            "contexts");
diff --git a/vp9/encoder/vp9_temporal_filter.c b/vp9/encoder/vp9_temporal_filter.c
index a4051f05e..424cc0843 100644
--- a/vp9/encoder/vp9_temporal_filter.c
+++ b/vp9/encoder/vp9_temporal_filter.c
@@ -710,8 +710,9 @@ void vp9_temporal_filter(VP9_COMP *cpi, int distance) {
 #if CONFIG_VP9_HIGHBITDEPTH
                                        cm->use_highbitdepth,
 #endif
-                                       VP9_ENC_BORDER_IN_PIXELS, NULL, NULL,
-                                       NULL)) {
+                                       VP9_ENC_BORDER_IN_PIXELS,
+                                       cm->byte_alignment,
+                                       NULL, NULL, NULL)) {
             vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                                "Failed to reallocate alt_ref_buffer");
           }
diff --git a/vp9/vp9_dx_iface.c b/vp9/vp9_dx_iface.c
index 809514001..43bf35f9c 100644
--- a/vp9/vp9_dx_iface.c
+++ b/vp9/vp9_dx_iface.c
@@ -44,6 +44,7 @@ struct vpx_codec_alg_priv {
   int                     flushed;
   int                     invert_tile_order;
   int                     frame_parallel_decode;  // frame-based threading.
+  int                     byte_alignment;
 
   // External frame buffer info to save for VP9 common.
   void *ext_priv;  // Private data associated with the external frame buffers.
@@ -219,6 +220,7 @@ static void init_buffer_callbacks(vpx_codec_alg_priv_t *ctx) {
   VP9_COMMON *const cm = &ctx->pbi->common;
 
   cm->new_fb_idx = -1;
+  cm->byte_alignment = ctx->byte_alignment;
 
   if (ctx->get_ext_fb_cb != NULL && ctx->release_ext_fb_cb != NULL) {
     cm->get_fb_cb = ctx->get_ext_fb_cb;
@@ -617,6 +619,27 @@ static vpx_codec_err_t ctrl_set_decryptor(vpx_codec_alg_priv_t *ctx,
   return VPX_CODEC_OK;
 }
 
+static vpx_codec_err_t ctrl_set_byte_alignment(vpx_codec_alg_priv_t *ctx,
+                                               va_list args) {
+  const int legacy_byte_alignment = 0;
+  const int min_byte_alignment = 32;
+  const int max_byte_alignment = 1024;
+  const int byte_alignment = va_arg(args, int);
+
+  if (byte_alignment != legacy_byte_alignment &&
+      (byte_alignment < min_byte_alignment ||
+       byte_alignment > max_byte_alignment ||
+       (byte_alignment & (byte_alignment - 1)) != 0))
+    return VPX_CODEC_INVALID_PARAM;
+
+  ctx->byte_alignment = byte_alignment;
+  if (ctx->pbi != NULL) {
+    VP9_COMMON *const cm = &ctx->pbi->common;
+    cm->byte_alignment = byte_alignment;
+  }
+  return VPX_CODEC_OK;
+}
+
 static vpx_codec_ctrl_fn_map_t decoder_ctrl_maps[] = {
   {VP8_COPY_REFERENCE,            ctrl_copy_reference},
 
@@ -629,6 +652,7 @@ static vpx_codec_ctrl_fn_map_t decoder_ctrl_maps[] = {
   {VP8_SET_DBG_DISPLAY_MV,        ctrl_set_dbg_options},
   {VP9_INVERT_TILE_DECODE_ORDER,  ctrl_set_invert_tile_order},
   {VPXD_SET_DECRYPTOR,            ctrl_set_decryptor},
+  {VP9_SET_BYTE_ALIGNMENT,        ctrl_set_byte_alignment},
 
   // Getters
   {VP8D_GET_LAST_REF_UPDATES,     ctrl_get_last_ref_updates},
diff --git a/vpx/vp8dx.h b/vpx/vp8dx.h
index 379b30620..5cc25cd6a 100644
--- a/vpx/vp8dx.h
+++ b/vpx/vp8dx.h
@@ -78,6 +78,13 @@ enum vp8_dec_control_id {
   /** control function to get the bit depth of the stream. */
   VP9D_GET_BIT_DEPTH,
 
+  /** control function to set the byte alignment of the planes in the reference
+   * buffers. Valid values are power of 2, from 32 to 1024. A value of 0 sets
+   * legacy alignment. I.e. Y plane is aligned to 32 bytes, U plane directly
+   * follows Y plane, and V plane directly follows U plane. Default value is 0.
+   */
+  VP9_SET_BYTE_ALIGNMENT,
+
   /** For testing. */
   VP9_INVERT_TILE_DECODE_ORDER,
 
diff --git a/vpx_scale/generic/yv12config.c b/vpx_scale/generic/yv12config.c
index 00a8c163a..ff49ffb95 100644
--- a/vpx_scale/generic/yv12config.c
+++ b/vpx_scale/generic/yv12config.c
@@ -143,22 +143,25 @@ int vp9_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
                              int use_highbitdepth,
 #endif
                              int border,
+                             int byte_alignment,
                              vpx_codec_frame_buffer_t *fb,
                              vpx_get_frame_buffer_cb_fn_t cb,
                              void *cb_priv) {
   if (ybf) {
+    const int vp9_byte_align = (byte_alignment == 0) ? 1 : byte_alignment;
     const int aligned_width = (width + 7) & ~7;
     const int aligned_height = (height + 7) & ~7;
     const int y_stride = ((aligned_width + 2 * border) + 31) & ~31;
     const uint64_t yplane_size = (aligned_height + 2 * border) *
-                                 (uint64_t)y_stride;
+                                 (uint64_t)y_stride + byte_alignment;
     const int uv_width = aligned_width >> ss_x;
     const int uv_height = aligned_height >> ss_y;
     const int uv_stride = y_stride >> ss_x;
     const int uv_border_w = border >> ss_x;
     const int uv_border_h = border >> ss_y;
     const uint64_t uvplane_size = (uv_height + 2 * uv_border_h) *
-                                  (uint64_t)uv_stride;
+                                  (uint64_t)uv_stride + byte_alignment;
+
 #if CONFIG_ALPHA
     const int alpha_width = aligned_width;
     const int alpha_height = aligned_height;
@@ -166,7 +169,7 @@ int vp9_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
     const int alpha_border_w = border;
     const int alpha_border_h = border;
     const uint64_t alpha_plane_size = (alpha_height + 2 * alpha_border_h) *
-                                      (uint64_t)alpha_stride;
+                                      (uint64_t)alpha_stride + byte_alignment;
 #if CONFIG_VP9_HIGHBITDEPTH
     const uint64_t frame_size = (1 + use_highbitdepth) *
         (yplane_size + 2 * uvplane_size + alpha_plane_size);
@@ -182,6 +185,9 @@ int vp9_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
     const uint64_t frame_size = yplane_size + 2 * uvplane_size;
 #endif  // CONFIG_VP9_HIGHBITDEPTH
 #endif  // CONFIG_ALPHA
+
+    uint8_t *buf = NULL;
+
     if (cb != NULL) {
       const int align_addr_extra_size = 31;
       const uint64_t external_frame_size = frame_size + align_addr_extra_size;
@@ -244,38 +250,33 @@ int vp9_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
     ybf->subsampling_x = ss_x;
     ybf->subsampling_y = ss_y;
 
+    buf = ybf->buffer_alloc;
 #if CONFIG_VP9_HIGHBITDEPTH
     if (use_highbitdepth) {
       // Store uint16 addresses when using 16bit framebuffers
-      uint8_t *p = CONVERT_TO_BYTEPTR(ybf->buffer_alloc);
-      ybf->y_buffer = p + (border * y_stride) + border;
-      ybf->u_buffer = p + yplane_size +
-          (uv_border_h * uv_stride) + uv_border_w;
-      ybf->v_buffer = p + yplane_size + uvplane_size +
-          (uv_border_h * uv_stride) + uv_border_w;
+      buf = CONVERT_TO_BYTEPTR(ybf->buffer_alloc);
       ybf->flags = YV12_FLAG_HIGHBITDEPTH;
     } else {
-      ybf->y_buffer = ybf->buffer_alloc + (border * y_stride) + border;
-      ybf->u_buffer = ybf->buffer_alloc + yplane_size +
-          (uv_border_h * uv_stride) + uv_border_w;
-      ybf->v_buffer = ybf->buffer_alloc + yplane_size + uvplane_size +
-          (uv_border_h * uv_stride) + uv_border_w;
       ybf->flags = 0;
     }
-#else
-    ybf->y_buffer = ybf->buffer_alloc + (border * y_stride) + border;
-    ybf->u_buffer = ybf->buffer_alloc + yplane_size +
-                    (uv_border_h * uv_stride) + uv_border_w;
-    ybf->v_buffer = ybf->buffer_alloc + yplane_size + uvplane_size +
-                    (uv_border_h * uv_stride) + uv_border_w;
 #endif  // CONFIG_VP9_HIGHBITDEPTH
 
+    ybf->y_buffer = (uint8_t *)yv12_align_addr(
+        buf + (border * y_stride) + border, vp9_byte_align);
+    ybf->u_buffer = (uint8_t *)yv12_align_addr(
+        buf + yplane_size + (uv_border_h * uv_stride) + uv_border_w,
+        vp9_byte_align);
+    ybf->v_buffer = (uint8_t *)yv12_align_addr(
+        buf + yplane_size + uvplane_size + (uv_border_h * uv_stride) +
+        uv_border_w, vp9_byte_align);
+
 #if CONFIG_ALPHA
     ybf->alpha_width = alpha_width;
     ybf->alpha_height = alpha_height;
     ybf->alpha_stride = alpha_stride;
-    ybf->alpha_buffer = ybf->buffer_alloc + yplane_size + 2 * uvplane_size +
-                        (alpha_border_h * alpha_stride) + alpha_border_w;
+    ybf->alpha_buffer = (uint8_t *)yv12_align_addr(
+        buf + yplane_size + 2 * uvplane_size +
+        (alpha_border_h * alpha_stride) + alpha_border_w, vp9_byte_align);
 #endif
     ybf->corrupted = 0; /* assume not corrupted by errors */
     return 0;
@@ -289,14 +290,15 @@ int vp9_alloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
 #if CONFIG_VP9_HIGHBITDEPTH
                            int use_highbitdepth,
 #endif
-                           int border) {
+                           int border,
+                           int byte_alignment) {
   if (ybf) {
     vp9_free_frame_buffer(ybf);
     return vp9_realloc_frame_buffer(ybf, width, height, ss_x, ss_y,
 #if CONFIG_VP9_HIGHBITDEPTH
                                     use_highbitdepth,
 #endif
-                                    border, NULL, NULL, NULL);
+                                    border, byte_alignment, NULL, NULL, NULL);
   }
   return -2;
 }
diff --git a/vpx_scale/yv12config.h b/vpx_scale/yv12config.h
index b9f13fd89..f04dee1e8 100644
--- a/vpx_scale/yv12config.h
+++ b/vpx_scale/yv12config.h
@@ -73,9 +73,10 @@ int vp9_alloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
 #if CONFIG_VP9_HIGHBITDEPTH
                            int use_highbitdepth,
 #endif
-                           int border);
+                           int border, int byte_alignment);
 
-// Updates the yv12 buffer config with the frame buffer. If cb is not
+// Updates the yv12 buffer config with the frame buffer. |byte_alignment| must
+// be a power of 2, from 32 to 1024. 0 sets legacy alignment. If cb is not
 // NULL, then libvpx is using the frame buffer callbacks to handle memory.
 // If cb is not NULL, libvpx will call cb with minimum size in bytes needed
 // to decode the current frame. If cb is NULL, libvpx will allocate memory
@@ -87,6 +88,7 @@ int vp9_realloc_frame_buffer(YV12_BUFFER_CONFIG *ybf,
                              int use_highbitdepth,
 #endif
                              int border,
+                             int byte_alignment,
                              vpx_codec_frame_buffer_t *fb,
                              vpx_get_frame_buffer_cb_fn_t cb,
                              void *cb_priv);