Add error handling for frame parallel decode and unit test for that.

Change-Id: I6e309e11f1641618d2424b7a2c0fe744b8974dec

test/codec_factory.h

@@ -35,6 +35,10 @@ class CodecFactory {
   virtual Decoder* CreateDecoder(vpx_codec_dec_cfg_t cfg,
                                  unsigned long deadline) const = 0;
 
+  virtual Decoder* CreateDecoder(vpx_codec_dec_cfg_t cfg,
+                                 const vpx_codec_flags_t flags,
+                                 unsigned long deadline) const = 0;  // NOLINT
+
   virtual Encoder* CreateEncoder(vpx_codec_enc_cfg_t cfg,
                                  unsigned long deadline,
                                  const unsigned long init_flags,
@@ -72,6 +76,10 @@ class VP8Decoder : public Decoder {
   VP8Decoder(vpx_codec_dec_cfg_t cfg, unsigned long deadline)
       : Decoder(cfg, deadline) {}
 
+  VP8Decoder(vpx_codec_dec_cfg_t cfg, const vpx_codec_flags_t flag,
+             unsigned long deadline)  // NOLINT
+      : Decoder(cfg, flag, deadline) {}
+
  protected:
   virtual vpx_codec_iface_t* CodecInterface() const {
 #if CONFIG_VP8_DECODER
@@ -104,8 +112,14 @@ class VP8CodecFactory : public CodecFactory {
 
   virtual Decoder* CreateDecoder(vpx_codec_dec_cfg_t cfg,
                                  unsigned long deadline) const {
+    return CreateDecoder(cfg, 0, deadline);
+  }
+
+  virtual Decoder* CreateDecoder(vpx_codec_dec_cfg_t cfg,
+                                 const vpx_codec_flags_t flags,
+                                 unsigned long deadline) const {  // NOLINT
 #if CONFIG_VP8_DECODER
-    return new VP8Decoder(cfg, deadline);
+    return new VP8Decoder(cfg, flags, deadline);
 #else
     return NULL;
 #endif
@@ -154,6 +168,10 @@ class VP9Decoder : public Decoder {
   VP9Decoder(vpx_codec_dec_cfg_t cfg, unsigned long deadline)
       : Decoder(cfg, deadline) {}
 
+  VP9Decoder(vpx_codec_dec_cfg_t cfg, const vpx_codec_flags_t flag,
+             unsigned long deadline)  // NOLINT
+      : Decoder(cfg, flag, deadline) {}
+
  protected:
   virtual vpx_codec_iface_t* CodecInterface() const {
 #if CONFIG_VP9_DECODER
@@ -186,8 +204,14 @@ class VP9CodecFactory : public CodecFactory {
 
   virtual Decoder* CreateDecoder(vpx_codec_dec_cfg_t cfg,
                                  unsigned long deadline) const {
+    return CreateDecoder(cfg, 0, deadline);
+  }
+
+  virtual Decoder* CreateDecoder(vpx_codec_dec_cfg_t cfg,
+                                 const vpx_codec_flags_t flags,
+                                 unsigned long deadline) const {  // NOLINT
 #if CONFIG_VP9_DECODER
-    return new VP9Decoder(cfg, deadline);
+    return new VP9Decoder(cfg, flags, deadline);
 #else
     return NULL;
 #endif

test/decode_test_driver.cc

@@ -40,26 +40,28 @@ vpx_codec_err_t Decoder::DecodeFrame(const uint8_t *cxdata, size_t size,
 }
 
 void DecoderTest::RunLoop(CompressedVideoSource *video) {
-  vpx_codec_dec_cfg_t dec_cfg = {0};
-  Decoder* const decoder = codec_->CreateDecoder(dec_cfg, 0);
+  Decoder* const decoder = codec_->CreateDecoder(cfg_, flags_, 0);
   ASSERT_TRUE(decoder != NULL);
   const char *codec_name = decoder->GetDecoderName();
   const bool is_vp8 = strncmp(kVP8Name, codec_name, sizeof(kVP8Name) - 1) == 0;
+  bool end_of_file = false;
 
   // Decode frames.
-  for (video->Begin(); !::testing::Test::HasFailure() && video->cxdata();
+  for (video->Begin(); !::testing::Test::HasFailure() && !end_of_file;
        video->Next()) {
     PreDecodeFrameHook(*video, decoder);
 
     vpx_codec_stream_info_t stream_info;
     stream_info.sz = sizeof(stream_info);
+
+    if (video->cxdata() != NULL) {
       const vpx_codec_err_t res_peek = decoder->PeekStream(video->cxdata(),
                                                            video->frame_size(),
                                                            &stream_info);
       if (is_vp8) {
         /* Vp8's implementation of PeekStream returns an error if the frame you
-       * pass it is not a keyframe, so we only expect VPX_CODEC_OK on the first
-       * frame, which must be a keyframe. */
+         * pass it is not a keyframe, so we only expect VPX_CODEC_OK on the
+         * first frame, which must be a keyframe. */
         if (video->frame_number() == 0)
           ASSERT_EQ(VPX_CODEC_OK, res_peek) << "Peek return failed: "
               << vpx_codec_err_to_string(res_peek);
@@ -74,6 +76,12 @@ void DecoderTest::RunLoop(CompressedVideoSource *video) {
                                                      video->frame_size());
       if (!HandleDecodeResult(res_dec, *video, decoder))
         break;
+    } else {
+      // Signal end of the file to the decoder.
+      const vpx_codec_err_t res_dec = decoder->DecodeFrame(NULL, 0);
+      ASSERT_EQ(VPX_CODEC_OK, res_dec) << decoder->DecodeError();
+      end_of_file = true;
+    }
 
     DxDataIterator dec_iter = decoder->GetDxData();
     const vpx_image_t *img = NULL;
@@ -85,4 +93,12 @@ void DecoderTest::RunLoop(CompressedVideoSource *video) {
 
   delete decoder;
 }
+
+void DecoderTest::set_cfg(const vpx_codec_dec_cfg_t &dec_cfg) {
+  memcpy(&cfg_, &dec_cfg, sizeof(cfg_));
+}
+
+void DecoderTest::set_flags(const vpx_codec_flags_t flags) {
+  flags_ = flags;
+}
 }  // namespace libvpx_test

test/decode_test_driver.h

@@ -41,7 +41,13 @@ class DxDataIterator {
 class Decoder {
  public:
   Decoder(vpx_codec_dec_cfg_t cfg, unsigned long deadline)
-      : cfg_(cfg), deadline_(deadline), init_done_(false) {
+      : cfg_(cfg), flags_(0), deadline_(deadline), init_done_(false) {
+    memset(&decoder_, 0, sizeof(decoder_));
+  }
+
+  Decoder(vpx_codec_dec_cfg_t cfg, const vpx_codec_flags_t flag,
+          unsigned long deadline)  // NOLINT
+      : cfg_(cfg), flags_(flag), deadline_(deadline), init_done_(false) {
     memset(&decoder_, 0, sizeof(decoder_));
   }
 
@@ -102,7 +108,7 @@ class Decoder {
     if (!init_done_) {
       const vpx_codec_err_t res = vpx_codec_dec_init(&decoder_,
                                                      CodecInterface(),
-                                                     &cfg_, 0);
+                                                     &cfg_, flags_);
       ASSERT_EQ(VPX_CODEC_OK, res) << DecodeError();
       init_done_ = true;
     }
@@ -110,6 +116,7 @@ class Decoder {
 
   vpx_codec_ctx_t     decoder_;
   vpx_codec_dec_cfg_t cfg_;
+  vpx_codec_flags_t   flags_;
   unsigned int        deadline_;
   bool                init_done_;
 };
@@ -120,6 +127,9 @@ class DecoderTest {
   // Main decoding loop
   virtual void RunLoop(CompressedVideoSource *video);
 
+  virtual void set_cfg(const vpx_codec_dec_cfg_t &dec_cfg);
+  virtual void set_flags(const vpx_codec_flags_t flags);
+
   // Hook to be called before decompressing every frame.
   virtual void PreDecodeFrameHook(const CompressedVideoSource& video,
                                   Decoder *decoder) {}
@@ -137,11 +147,15 @@ class DecoderTest {
                                      const unsigned int frame_number) {}
 
  protected:
-  explicit DecoderTest(const CodecFactory *codec) : codec_(codec) {}
+  explicit DecoderTest(const CodecFactory *codec) : codec_(codec), flags_(0) {
+    memset(&cfg_, 0, sizeof(cfg_));
+  }
 
   virtual ~DecoderTest() {}
 
   const CodecFactory *codec_;
+  vpx_codec_dec_cfg_t cfg_;
+  vpx_codec_flags_t   flags_;
 };
 
 }  // namespace libvpx_test

test/invalid_file_test.cc

@@ -97,7 +97,7 @@ const char *const kVP9InvalidFileTests[] = {
   "invalid-vp90-01.webm",
   "invalid-vp90-02.webm",
   "invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.ivf",
-  "invalid-vp90-03.webm",
+  "invalid-vp90-03-v3.webm",
   "invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf",
   "invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf",
 };

test/test-data.sha1

@@ -4,8 +4,8 @@ fe346136b9b8c1e6f6084cc106485706915795e4  invalid-vp90-01.webm
 25751f5d3b05ff03f0719ad42cd625348eb8961e  invalid-vp90-01.webm.res
 d78e2fceba5ac942246503ec8366f879c4775ca5  invalid-vp90-02.webm
 2dadee5306245fa5eeb0f99652d0e17afbcba96d  invalid-vp90-02.webm.res
-df1a1453feb3c00d7d89746c7003b4163523bff3  invalid-vp90-03.webm
-8fe6fd82bf537340f586f97a7ae31fb37ccda302  invalid-vp90-03.webm.res
+df1a1453feb3c00d7d89746c7003b4163523bff3  invalid-vp90-03-v3.webm
+4935c62becc68c13642a03db1e6d3e2331c1c612  invalid-vp90-03-v3.webm.res
 a432f96ff0a787268e2f94a8092ab161a18d1b06  park_joy_90p_10_420.y4m
 0b194cc312c3a2e84d156a221b0a5eb615dfddc5  park_joy_90p_10_422.y4m
 ff0e0a21dc2adc95b8c1b37902713700655ced17  park_joy_90p_10_444.y4m
@@ -661,3 +661,12 @@ d3964f9dad9f60363c81b688324d95b4ec7c8038  invalid-vp90-2-00-quantizer-00.webm.iv
 456d1493e52d32a5c30edf44a27debc1fa6b253a  invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf.res
 c123d1f9f02fb4143abb5e271916e3a3080de8f6  invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf
 456d1493e52d32a5c30edf44a27debc1fa6b253a  invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf.res
+f97088c7359fc8d3d5aa5eafe57bc7308b3ee124  vp90-2-20-big_superframe-01.webm
+47d7d409785afa33b123376de0c907336e6c7bd7  vp90-2-20-big_superframe-01.webm.md5
+65ade6d2786209582c50d34cfe22b3cdb033abaf  vp90-2-20-big_superframe-02.webm
+7c0ed8d04c4d06c5411dd2e5de2411d37f092db5  vp90-2-20-big_superframe-02.webm.md5
+667ec8718c982aef6be07eb94f083c2efb9d2d16  vp90-2-07-frame_parallel-1.webm
+bfc82bf848e9c05020d61e3ffc1e62f25df81d19  vp90-2-07-frame_parallel-1.webm.md5
+efd5a51d175cfdacd169ed23477729dc558030dc  invalid-vp90-2-07-frame_parallel-1.webm
+9f912712ec418be69adb910e2ca886a63c4cec08  invalid-vp90-2-07-frame_parallel-2.webm
+445f5a53ca9555341852997ccdd480a51540bd14  invalid-vp90-2-07-frame_parallel-3.webm

test/test.mk

@@ -31,6 +31,7 @@ LIBVPX_TEST_SRCS-$(CONFIG_VP8_ENCODER) += keyframe_test.cc
 
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += external_frame_buffer_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += user_priv_test.cc
+LIBVPX_TEST_SRCS-$(CONFIG_VP9_DECODER) += vp9_frame_parallel_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += active_map_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += borders_test.cc
 LIBVPX_TEST_SRCS-$(CONFIG_VP9_ENCODER) += cpu_speed_test.cc
@@ -683,6 +684,8 @@ LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-06-bilinear.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-06-bilinear.webm.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-07-frame_parallel.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-07-frame_parallel.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-07-frame_parallel-1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-07-frame_parallel-1.webm.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile-4x1.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile-4x1.webm.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-08-tile-4x4.webm
@@ -773,20 +776,27 @@ LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-15-segkey_adpq.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-15-segkey_adpq.webm.md5
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yuv444.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp91-2-04-yuv444.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-20-big_superframe-01.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-20-big_superframe-01.webm.md5
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-20-big_superframe-02.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += vp90-2-20-big_superframe-02.webm.md5
 
 # Invalid files for testing libvpx error checking.
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-01.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-01.webm.res
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-02.webm
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-02.webm.res
-LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-03.webm
-LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-03.webm.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-03-v3.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-03-v3.webm.res
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.ivf
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-00.webm.ivf.s5861_r01-05_b6-.ivf.res
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-.ivf.res
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf
 LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-00-quantizer-11.webm.ivf.s52984_r01-05_b6-z.ivf.res
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-07-frame_parallel-1.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-07-frame_parallel-2.webm
+LIBVPX_TEST_DATA-$(CONFIG_VP9_DECODER) += invalid-vp90-2-07-frame_parallel-3.webm
 
 ifeq ($(CONFIG_DECODE_PERF_TESTS),yes)
 # BBB VP9 streams

test/test_vector_test.cc

@@ -12,6 +12,7 @@
 #include <cstdlib>
 #include <string>
 #include "third_party/googletest/src/include/gtest/gtest.h"
+#include "../tools_common.h"
 #include "./vpx_config.h"
 #include "test/codec_factory.h"
 #include "test/decode_test_driver.h"
@@ -26,10 +27,24 @@
 
 namespace {
 
+enum DecodeMode {
+  kSerialMode,
+  kFrameParallMode
+};
+
+const int kDecodeMode = 0;
+const int kThreads = 1;
+const int kFileName = 2;
+
+typedef std::tr1::tuple<int, int, const char *> DecodeParam;
+
 class TestVectorTest : public ::libvpx_test::DecoderTest,
-    public ::libvpx_test::CodecTestWithParam<const char*> {
+    public ::libvpx_test::CodecTestWithParam<DecodeParam> {
  protected:
-  TestVectorTest() : DecoderTest(GET_PARAM(0)), md5_file_(NULL) {}
+  TestVectorTest()
+      : DecoderTest(GET_PARAM(0)),
+        md5_file_(NULL) {
+  }
 
   virtual ~TestVectorTest() {
     if (md5_file_)
@@ -71,8 +86,25 @@ class TestVectorTest : public ::libvpx_test::DecoderTest,
 // checksums match the correct md5 data, then the test is passed. Otherwise,
 // the test failed.
 TEST_P(TestVectorTest, MD5Match) {
-  const std::string filename = GET_PARAM(1);
+  const DecodeParam input = GET_PARAM(1);
+  const std::string filename = std::tr1::get<kFileName>(input);
+  const int threads = std::tr1::get<kThreads>(input);
+  const int mode = std::tr1::get<kDecodeMode>(input);
   libvpx_test::CompressedVideoSource *video = NULL;
+  vpx_codec_flags_t flags = 0;
+  vpx_codec_dec_cfg_t cfg = {0};
+  char str[256];
+
+  if (mode == kFrameParallMode) {
+    flags |= VPX_CODEC_USE_FRAME_THREADING;
+  }
+
+  cfg.threads = threads;
+
+  snprintf(str, sizeof(str) / sizeof(str[0]) - 1,
+           "file: %s  mode: %s threads: %d",
+           filename.c_str(), mode == 0 ? "Serial" : "Parallel", threads);
+  SCOPED_TRACE(str);
 
   // Open compressed video file.
   if (filename.substr(filename.length() - 3, 3) == "ivf") {
@@ -92,18 +124,53 @@ TEST_P(TestVectorTest, MD5Match) {
   const std::string md5_filename = filename + ".md5";
   OpenMD5File(md5_filename);
 
+  // Set decode config and flags.
+  set_cfg(cfg);
+  set_flags(flags);
+
   // Decode frame, and check the md5 matching.
   ASSERT_NO_FATAL_FAILURE(RunLoop(video));
   delete video;
 }
 
-VP8_INSTANTIATE_TEST_CASE(TestVectorTest,
+// Test VP8 decode in serial mode with single thread.
+// NOTE: VP8 only support serial mode.
+INSTANTIATE_TEST_CASE_P(
+    VP8, TestVectorTest,
+    ::testing::Combine(
+        ::testing::Values(
+            static_cast<const libvpx_test::CodecFactory *>(&libvpx_test::kVP8)),
+        ::testing::Combine(
+            ::testing::Values(0),  // Serial Mode.
+            ::testing::Values(1),  // Single thread.
             ::testing::ValuesIn(libvpx_test::kVP8TestVectors,
                                 libvpx_test::kVP8TestVectors +
-                                              libvpx_test::kNumVP8TestVectors));
-VP9_INSTANTIATE_TEST_CASE(TestVectorTest,
+                                    libvpx_test::kNumVP8TestVectors))));
+
+// Test VP9 decode in serial mode with single thread.
+INSTANTIATE_TEST_CASE_P(
+    VP9, TestVectorTest,
+    ::testing::Combine(
+        ::testing::Values(
+            static_cast<const libvpx_test::CodecFactory *>(&libvpx_test::kVP9)),
+        ::testing::Combine(
+            ::testing::Values(0),  // Serial Mode.
+            ::testing::Values(1),  // Single thread.
             ::testing::ValuesIn(libvpx_test::kVP9TestVectors,
                                 libvpx_test::kVP9TestVectors +
-                                              libvpx_test::kNumVP9TestVectors));
+                                    libvpx_test::kNumVP9TestVectors))));
 
+
+// Test VP9 decode in frame parallel mode with different number of threads.
+INSTANTIATE_TEST_CASE_P(
+    VP9MultiThreadedFrameParallel, TestVectorTest,
+    ::testing::Combine(
+        ::testing::Values(
+            static_cast<const libvpx_test::CodecFactory *>(&libvpx_test::kVP9)),
+        ::testing::Combine(
+            ::testing::Values(1),        // Frame Parallel mode.
+            ::testing::Range(2, 9),      // With 2 ~ 8 threads.
+            ::testing::ValuesIn(libvpx_test::kVP9TestVectors,
+                                libvpx_test::kVP9TestVectors +
+                                    libvpx_test::kNumVP9TestVectors))));
 }  // namespace

test/test_vectors.cc

@@ -154,7 +154,8 @@ const char *const kVP9TestVectors[] = {
   "vp90-2-03-size-226x210.webm", "vp90-2-03-size-226x224.webm",
   "vp90-2-03-size-226x226.webm", "vp90-2-03-deltaq.webm",
   "vp90-2-05-resize.ivf", "vp90-2-06-bilinear.webm",
-  "vp90-2-07-frame_parallel.webm", "vp90-2-08-tile_1x2_frame_parallel.webm",
+  "vp90-2-07-frame_parallel.webm", "vp90-2-07-frame_parallel-1.webm",
+  "vp90-2-08-tile_1x2_frame_parallel.webm",
   "vp90-2-08-tile_1x2.webm", "vp90-2-08-tile_1x4_frame_parallel.webm",
   "vp90-2-08-tile_1x4.webm", "vp90-2-08-tile_1x8_frame_parallel.webm",
   "vp90-2-08-tile_1x8.webm", "vp90-2-08-tile-4x4.webm",
@@ -181,6 +182,7 @@ const char *const kVP9TestVectors[] = {
   "vp90-2-14-resize-fp-tiles-8-2.webm", "vp90-2-14-resize-fp-tiles-8-4.webm",
   "vp90-2-15-segkey.webm", "vp90-2-15-segkey_adpq.webm",
   "vp91-2-04-yuv444.webm",
+  "vp90-2-20-big_superframe-01.webm", "vp90-2-20-big_superframe-02.webm",
 };
 const int kNumVP9TestVectors = NELEMENTS(kVP9TestVectors);
 #endif  // CONFIG_VP9_DECODER

test/vp9_frame_parallel_test.cc

@@ -0,0 +1,208 @@
+/*
+ *  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 <cstdio>
+#include <cstdlib>
+#include <string>
+#include "third_party/googletest/src/include/gtest/gtest.h"
+#include "./vpx_config.h"
+#include "test/codec_factory.h"
+#include "test/decode_test_driver.h"
+#include "test/ivf_video_source.h"
+#include "test/md5_helper.h"
+#include "test/util.h"
+#if CONFIG_WEBM_IO
+#include "test/webm_video_source.h"
+#endif
+#include "vpx_mem/vpx_mem.h"
+
+namespace {
+
+using std::string;
+
+#if CONFIG_WEBM_IO
+
+struct FileList {
+  const char *name;
+  // md5 sum for decoded frames which does not include skipped frames.
+  const char *expected_md5;
+  const int pause_frame_num;
+};
+
+// Decodes |filename| with |num_threads|. Pause at the specified frame_num,
+// seek to next key frame and then continue decoding until the end. Return
+// the md5 of the decoded frames which does not include skipped frames.
+string DecodeFile(const string &filename, int num_threads, int pause_num) {
+  libvpx_test::WebMVideoSource video(filename);
+  video.Init();
+  int in_frames = 0;
+  int out_frames = 0;
+
+  vpx_codec_dec_cfg_t cfg = {0};
+  cfg.threads = num_threads;
+  vpx_codec_flags_t flags = 0;
+  flags |= VPX_CODEC_USE_FRAME_THREADING;
+  libvpx_test::VP9Decoder decoder(cfg, flags, 0);
+
+  libvpx_test::MD5 md5;
+  video.Begin();
+
+  do {
+    ++in_frames;
+    const vpx_codec_err_t res =
+        decoder.DecodeFrame(video.cxdata(), video.frame_size());
+    if (res != VPX_CODEC_OK) {
+      EXPECT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+      break;
+    }
+
+    // Pause at specified frame number.
+    if (in_frames == pause_num) {
+      // Flush the decoder and then seek to next key frame.
+      decoder.DecodeFrame(NULL, 0);
+      video.SeekToNextKeyFrame();
+    } else {
+      video.Next();
+    }
+
+    // Flush the decoder at the end of the video.
+    if (!video.cxdata())
+      decoder.DecodeFrame(NULL, 0);
+
+    libvpx_test::DxDataIterator dec_iter = decoder.GetDxData();
+    const vpx_image_t *img;
+
+    // Get decompressed data
+    while ((img = dec_iter.Next())) {
+      ++out_frames;
+      md5.Add(img);
+    }
+  } while (video.cxdata() != NULL);
+
+  EXPECT_EQ(in_frames, out_frames) <<
+      "Input frame count does not match output frame count";
+
+  return string(md5.Get());
+}
+
+void DecodeFiles(const FileList files[]) {
+  for (const FileList *iter = files; iter->name != NULL; ++iter) {
+    SCOPED_TRACE(iter->name);
+    for (int t = 2; t <= 8; ++t) {
+      EXPECT_EQ(iter->expected_md5,
+                DecodeFile(iter->name, t, iter->pause_frame_num))
+          << "threads = " << t;
+    }
+  }
+}
+
+TEST(VP9MultiThreadedFrameParallel, PauseSeekResume) {
+  // vp90-2-07-frame_parallel-1.webm is a 40 frame video file with
+  // one key frame for every ten frames.
+  static const FileList files[] = {
+    { "vp90-2-07-frame_parallel-1.webm",
+      "6ea7c3875d67252e7caf2bc6e75b36b1", 6},
+    { "vp90-2-07-frame_parallel-1.webm",
+      "4bb634160c7356a8d7d4299b6dc83a45", 12},
+    { "vp90-2-07-frame_parallel-1.webm",
+      "89772591e6ef461f9fa754f916c78ed8", 26},
+    { NULL, NULL, 0},
+  };
+  DecodeFiles(files);
+}
+
+struct InvalidFileList {
+  const char *name;
+  // md5 sum for decoded frames which does not include corrupted frames.
+  const char *expected_md5;
+  // Expected number of decoded frames which does not include corrupted frames.
+  const int expected_frame_count;
+};
+
+// Decodes |filename| with |num_threads|. Return the md5 of the decoded
+// frames which does not include corrupted frames.
+string DecodeInvalidFile(const string &filename, int num_threads,
+                         int expected_frame_count) {
+  libvpx_test::WebMVideoSource video(filename);
+  video.Init();
+
+  vpx_codec_dec_cfg_t cfg = vpx_codec_dec_cfg_t();
+  cfg.threads = num_threads;
+  const vpx_codec_flags_t flags = VPX_CODEC_USE_FRAME_THREADING;
+  libvpx_test::VP9Decoder decoder(cfg, flags, 0);
+
+  libvpx_test::MD5 md5;
+  video.Begin();
+
+  int out_frames = 0;
+  do {
+    const vpx_codec_err_t res =
+        decoder.DecodeFrame(video.cxdata(), video.frame_size());
+    // TODO(hkuang): frame parallel mode should return an error on corruption.
+    if (res != VPX_CODEC_OK) {
+      EXPECT_EQ(VPX_CODEC_OK, res) << decoder.DecodeError();
+      break;
+    }
+
+    video.Next();
+
+    // Flush the decoder at the end of the video.
+    if (!video.cxdata())
+      decoder.DecodeFrame(NULL, 0);
+
+    libvpx_test::DxDataIterator dec_iter = decoder.GetDxData();
+    const vpx_image_t *img;
+
+    // Get decompressed data
+    while ((img = dec_iter.Next())) {
+      ++out_frames;
+      md5.Add(img);
+    }
+  } while (video.cxdata() != NULL);
+
+  EXPECT_EQ(expected_frame_count, out_frames) <<
+      "Input frame count does not match expected output frame count";
+
+  return string(md5.Get());
+}
+
+void DecodeInvalidFiles(const InvalidFileList files[]) {
+  for (const InvalidFileList *iter = files; iter->name != NULL; ++iter) {
+    SCOPED_TRACE(iter->name);
+    for (int t = 2; t <= 8; ++t) {
+      EXPECT_EQ(iter->expected_md5,
+                DecodeInvalidFile(iter->name, t, iter->expected_frame_count))
+          << "threads = " << t;
+    }
+  }
+}
+
+TEST(VP9MultiThreadedFrameParallel, InvalidFileTest) {
+  static const InvalidFileList files[] = {
+    // invalid-vp90-2-07-frame_parallel-1.webm is a 40 frame video file with
+    // one key frame for every ten frames. The 11th frame has corrupted data.
+    { "invalid-vp90-2-07-frame_parallel-1.webm",
+      "0549d0f45f60deaef8eb708e6c0eb6cb", 30},
+    // invalid-vp90-2-07-frame_parallel-2.webm is a 40 frame video file with
+    // one key frame for every ten frames. The 1st and 31st frames have
+    // corrupted data.
+    { "invalid-vp90-2-07-frame_parallel-2.webm",
+      "6a1f3cf6f9e7a364212fadb9580d525e", 20},
+    // invalid-vp90-2-07-frame_parallel-3.webm is a 40 frame video file with
+    // one key frame for every ten frames. The 13th frame has corrupted data.
+    { "invalid-vp90-2-07-frame_parallel-3.webm",
+      "a567c8259d27ad32b1b7f58db5ac89dd", 32},
+    { NULL, NULL, 0},
+  };
+  DecodeInvalidFiles(files);
+}
+
+#endif  // CONFIG_WEBM_IO
+}  // namespace

test/vp9_thread_test.cc

@@ -18,7 +18,7 @@
 #if CONFIG_WEBM_IO
 #include "test/webm_video_source.h"
 #endif
-#include "vp9/decoder/vp9_thread.h"
+#include "vp9/common/vp9_thread.h"
 
 namespace {
 
@@ -175,7 +175,7 @@ int Reset(VP9Worker *const /*worker*/) { return 1; }
 int Sync(VP9Worker *const worker) { return !worker->had_error; }
 
 void Execute(VP9Worker *const worker) {
-  worker->had_error |= worker->hook(worker->data1, worker->data2);
+  worker->had_error |= !worker->hook(worker->data1, worker->data2);
 }
 
 void Launch(VP9Worker *const worker) { Execute(worker); }

test/webm_video_source.h

@@ -69,6 +69,18 @@ class WebMVideoSource : public CompressedVideoSource {
     }
   }
 
+  void SeekToNextKeyFrame() {
+    ASSERT_TRUE(vpx_ctx_->file != NULL);
+    do {
+      const int status = webm_read_frame(webm_ctx_, &buf_, &buf_sz_, &buf_sz_);
+      ASSERT_GE(status, 0) << "webm_read_frame failed";
+      ++frame_;
+      if (status == 1) {
+        end_of_file_ = true;
+      }
+    } while (!webm_ctx_->is_key_frame && !end_of_file_);
+  }
+
   virtual const uint8_t *cxdata() const {
     return end_of_file_ ? NULL : buf_;
   }

vp8/vp8_dx_iface.c

@@ -60,6 +60,7 @@ struct vpx_codec_alg_priv
     vpx_decrypt_cb          decrypt_cb;
     void                    *decrypt_state;
     vpx_image_t             img;
+    int                     flushed;
     int                     img_setup;
     struct frame_buffers    yv12_frame_buffers;
     void                    *user_priv;
@@ -89,6 +90,7 @@ static void vp8_init_ctx(vpx_codec_ctx_t *ctx)
     ctx->priv->alg_priv->decrypt_cb = NULL;
     ctx->priv->alg_priv->decrypt_state = NULL;
     ctx->priv->init_flags = ctx->init_flags;
+    ctx->priv->alg_priv->flushed = 0;
 
     if (ctx->config.dec)
     {
@@ -327,6 +329,13 @@ static vpx_codec_err_t vp8_decode(vpx_codec_alg_priv_t  *ctx,
     unsigned int resolution_change = 0;
     unsigned int w, h;
 
+    if (data == NULL && data_sz == 0) {
+      ctx->flushed = 1;
+      return VPX_CODEC_OK;
+    }
+
+    /* Reset flushed when receiving a valid frame */
+    ctx->flushed = 0;
 
     /* Update the input fragment data */
     if(update_fragments(ctx, data, data_sz, &res) <= 0)

vp9/common/vp9_alloccommon.c

@@ -12,11 +12,37 @@
 #include "vpx_mem/vpx_mem.h"
 
 #include "vp9/common/vp9_blockd.h"
+#include "vp9/common/vp9_common.h"
 #include "vp9/common/vp9_entropymode.h"
 #include "vp9/common/vp9_entropymv.h"
 #include "vp9/common/vp9_onyxc_int.h"
 #include "vp9/common/vp9_systemdependent.h"
 
+// TODO(hkuang): Don't need to lock the whole pool after implementing atomic
+// frame reference count.
+void lock_buffer_pool(BufferPool *const pool) {
+#if CONFIG_MULTITHREAD
+  pthread_mutex_lock(&pool->pool_mutex);
+#else
+  (void)pool;
+#endif
+}
+
+void unlock_buffer_pool(BufferPool *const pool) {
+#if CONFIG_MULTITHREAD
+  pthread_mutex_unlock(&pool->pool_mutex);
+#else
+  (void)pool;
+#endif
+}
+
+static INLINE void alloc_mi_array(VP9_COMMON *cm, int mi_size, int idx) {
+  CHECK_MEM_ERROR(cm, cm->mip_array[idx],
+                  vpx_calloc(mi_size, sizeof(*cm->mip_array[0])));
+  CHECK_MEM_ERROR(cm, cm->mi_grid_base_array[idx],
+                  vpx_calloc(mi_size, sizeof(*cm->mi_grid_base_array[0])));
+}
+
 static void clear_mi_border(const VP9_COMMON *cm, MODE_INFO *mi) {
   int i;
 
@@ -49,40 +75,47 @@ static void setup_mi(VP9_COMMON *cm) {
   vpx_memset(cm->mi_grid_base, 0, cm->mi_stride * (cm->mi_rows + 1) *
                                       sizeof(*cm->mi_grid_base));
 
+  // Only clear mi border in non frame-parallel decode. In frame-parallel
+  // decode, prev_mip is managed by previous decoding thread. While in
+  // non frame-parallel decode, prev_mip and mip are both managed by
+  // current decoding thread.
+  if (!cm->frame_parallel_decode)
     clear_mi_border(cm, cm->prev_mip);
 }
 
 static int alloc_mi(VP9_COMMON *cm, int mi_size) {
   int i;
 
-  for (i = 0; i < 2; ++i) {
-    cm->mip_array[i] =
-        (MODE_INFO *)vpx_calloc(mi_size, sizeof(*cm->mip));
-    if (cm->mip_array[i] == NULL)
-      return 1;
-
-    cm->mi_grid_base_array[i] =
-        (MODE_INFO **)vpx_calloc(mi_size, sizeof(*cm->mi_grid_base));
-    if (cm->mi_grid_base_array[i] == NULL)
-      return 1;
+  for (i = 0; i < NUM_PING_PONG_BUFFERS; ++i) {
+    // Delay reallocation as another thread is accessing prev_mi.
+    if (cm->frame_parallel_decode && i == cm->prev_mi_idx) {
+      cm->update_prev_mi = 1;
+      continue;
+    }
+    alloc_mi_array(cm, mi_size, i);
   }
 
-  // Init the index.
+  cm->mip = cm->mip_array[cm->mi_idx];
+  cm->mi_grid_base = cm->mi_grid_base_array[cm->mi_idx];
+
+  if (!cm->frame_parallel_decode) {
     cm->mi_idx = 0;
     cm->prev_mi_idx = 1;
-
-  cm->mip = cm->mip_array[cm->mi_idx];
+    // In frame-parallel decode, prev_mip comes from another thread,
+    // so current decoding thread should not touch it.
     cm->prev_mip = cm->mip_array[cm->prev_mi_idx];
-  cm->mi_grid_base = cm->mi_grid_base_array[cm->mi_idx];
     cm->prev_mi_grid_base = cm->mi_grid_base_array[cm->prev_mi_idx];
+  }
 
   return 0;
 }
 
-static void free_mi(VP9_COMMON *cm) {
+static void free_mi(VP9_COMMON *cm, int decode_done) {
   int i;
 
-  for (i = 0; i < 2; ++i) {
+  for (i = 0; i < NUM_PING_PONG_BUFFERS; ++i) {
+    if (cm->frame_parallel_decode && i == cm->prev_mi_idx && !decode_done)
+      continue;
     vpx_free(cm->mip_array[i]);
     cm->mip_array[i] = NULL;
     vpx_free(cm->mi_grid_base_array[i]);
@@ -90,21 +123,62 @@ static void free_mi(VP9_COMMON *cm) {
   }
 
   cm->mip = NULL;
-  cm->prev_mip = NULL;
   cm->mi_grid_base = NULL;
+
+  if (!cm->frame_parallel_decode) {
+    cm->prev_mip = NULL;
     cm->prev_mi_grid_base = NULL;
+  }
+}
+
+static int alloc_seg_map(VP9_COMMON *cm, int seg_map_size) {
+  int i;
+
+  for (i = 0; i < NUM_PING_PONG_BUFFERS; ++i) {
+    cm->seg_map_array[i] = (uint8_t *)vpx_calloc(seg_map_size, 1);
+    if (cm->seg_map_array[i] == NULL)
+      return 1;
+  }
+
+  // Init the index.
+  cm->seg_map_idx = 0;
+  cm->prev_seg_map_idx = 1;
+
+  cm->current_frame_seg_map = cm->seg_map_array[cm->seg_map_idx];
+
+  if (!cm->frame_parallel_decode) {
+    cm->last_frame_seg_map = cm->seg_map_array[cm->prev_seg_map_idx];
+  }
+
+  return 0;
+}
+
+static void free_seg_map(VP9_COMMON *cm) {
+  int i;
+
+  for (i = 0; i < NUM_PING_PONG_BUFFERS; ++i) {
+    vpx_free(cm->seg_map_array[i]);
+    cm->seg_map_array[i] = NULL;
+  }
+
+  cm->current_frame_seg_map = NULL;
+
+  if (!cm->frame_parallel_decode) {
+    cm->last_frame_seg_map = NULL;
+  }
 }
 
 void vp9_free_frame_buffers(VP9_COMMON *cm) {
   int i;
+  BufferPool *const pool = cm->buffer_pool;
 
   for (i = 0; i < FRAME_BUFFERS; ++i) {
-    vp9_free_frame_buffer(&cm->frame_bufs[i].buf);
+    vp9_free_frame_buffer(&pool->frame_bufs[i].buf);
 
-    if (cm->frame_bufs[i].ref_count > 0 &&
-        cm->frame_bufs[i].raw_frame_buffer.data != NULL) {
-      cm->release_fb_cb(cm->cb_priv, &cm->frame_bufs[i].raw_frame_buffer);
-      cm->frame_bufs[i].ref_count = 0;
+    if (pool->frame_bufs[i].ref_count > 0 &&
+        pool->frame_bufs[i].raw_frame_buffer.data != NULL) {
+      pool->release_fb_cb(pool->cb_priv, &pool->frame_bufs[i].raw_frame_buffer);
+      pool->frame_bufs[i].ref_count = 0;
     }
   }
 
@@ -112,10 +186,8 @@ void vp9_free_frame_buffers(VP9_COMMON *cm) {
 }
 
 void vp9_free_context_buffers(VP9_COMMON *cm) {
-  free_mi(cm);
-
-  vpx_free(cm->last_frame_seg_map);
-  cm->last_frame_seg_map = NULL;
+  free_mi(cm, 1);
+  free_seg_map(cm);
 
   vpx_free(cm->above_context);
   cm->above_context = NULL;
@@ -139,16 +211,15 @@ int vp9_resize_frame_buffers(VP9_COMMON *cm, int width, int height) {
 
   set_mb_mi(cm, aligned_width, aligned_height);
 
-  free_mi(cm);
+  free_mi(cm, 0);
   if (alloc_mi(cm, cm->mi_stride * (cm->mi_rows + MI_BLOCK_SIZE)))
     goto fail;
 
   setup_mi(cm);
 
   // Create the segmentation map structure and set to 0.
-  vpx_free(cm->last_frame_seg_map);
-  cm->last_frame_seg_map = (uint8_t *)vpx_calloc(cm->mi_rows * cm->mi_cols, 1);
-  if (!cm->last_frame_seg_map)
+  free_seg_map(cm);
+  if (alloc_seg_map(cm, cm->mi_rows * cm->mi_cols))
     goto fail;
 
   vpx_free(cm->above_context);
@@ -176,13 +247,14 @@ int vp9_resize_frame_buffers(VP9_COMMON *cm, int width, int height) {
 
 static void init_frame_bufs(VP9_COMMON *cm) {
   int i;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
 
   cm->new_fb_idx = FRAME_BUFFERS - 1;
-  cm->frame_bufs[cm->new_fb_idx].ref_count = 1;
+  frame_bufs[cm->new_fb_idx].ref_count = 1;
 
   for (i = 0; i < REF_FRAMES; ++i) {
     cm->ref_frame_map[i] = i;
-    cm->frame_bufs[i].ref_count = 1;
+    frame_bufs[i].ref_count = 1;
   }
 }
 
@@ -190,12 +262,13 @@ int vp9_alloc_frame_buffers(VP9_COMMON *cm, int width, int height) {
   int i;
   const int ss_x = cm->subsampling_x;
   const int ss_y = cm->subsampling_y;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
 
   vp9_free_frame_buffers(cm);
 
   for (i = 0; i < FRAME_BUFFERS; ++i) {
-    cm->frame_bufs[i].ref_count = 0;
-    if (vp9_alloc_frame_buffer(&cm->frame_bufs[i].buf, width, height,
+    frame_bufs[i].ref_count = 0;
+    if (vp9_alloc_frame_buffer(&frame_bufs[i].buf, width, height,
                                ss_x, ss_y, VP9_ENC_BORDER_IN_PIXELS) < 0)
       goto fail;
   }
@@ -256,7 +329,6 @@ int vp9_alloc_context_buffers(VP9_COMMON *cm, int width, int height) {
 void vp9_remove_common(VP9_COMMON *cm) {
   vp9_free_frame_buffers(cm);
   vp9_free_context_buffers(cm);
-  vp9_free_internal_frame_buffers(&cm->int_frame_buffers);
 }
 
 void vp9_update_frame_size(VP9_COMMON *cm) {
@@ -267,13 +339,27 @@ void vp9_update_frame_size(VP9_COMMON *cm) {
   setup_mi(cm);
 
   // Initialize the previous frame segment map to 0.
-  if (cm->last_frame_seg_map)
-    vpx_memset(cm->last_frame_seg_map, 0, cm->mi_rows * cm->mi_cols);
+  if (cm->current_frame_seg_map)
+    vpx_memset(cm->current_frame_seg_map, 0, cm->mi_rows * cm->mi_cols);
 }
 
 void vp9_swap_mi_and_prev_mi(VP9_COMMON *cm) {
   // Swap indices.
   const int tmp = cm->mi_idx;
+
+  // Only used in frame parallel decode: Update the prev_mi buffer if
+  // needed. The worker that was accessing it must already finish decoding.
+  // So it can be resized safely now.
+  if (cm->update_prev_mi) {
+    const int mi_size = cm->mi_stride * (cm->mi_rows + MI_BLOCK_SIZE);
+    vpx_free(cm->mip_array[cm->prev_mi_idx]);
+    vpx_free(cm->mi_grid_base_array[cm->prev_mi_idx]);
+    cm->mip_array[cm->prev_mi_idx] = NULL;
+    cm->mi_grid_base_array[cm->prev_mi_idx] = NULL;
+    alloc_mi_array(cm, mi_size, cm->prev_mi_idx);
+    cm->update_prev_mi = 0;
+  }
+
   cm->mi_idx = cm->prev_mi_idx;
   cm->prev_mi_idx = tmp;
 
@@ -289,3 +375,13 @@ void vp9_swap_mi_and_prev_mi(VP9_COMMON *cm) {
   cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
   cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
 }
+
+void vp9_swap_current_and_last_seg_map(VP9_COMMON *cm) {
+  // Swap indices.
+  const int tmp = cm->seg_map_idx;
+  cm->seg_map_idx = cm->prev_seg_map_idx;
+  cm->prev_seg_map_idx = tmp;
+
+  cm->current_frame_seg_map = cm->seg_map_array[cm->seg_map_idx];
+  cm->last_frame_seg_map = cm->seg_map_array[cm->prev_seg_map_idx];
+}

vp9/common/vp9_alloccommon.h

@@ -34,6 +34,8 @@ void vp9_update_frame_size(struct VP9Common *cm);
 
 void vp9_swap_mi_and_prev_mi(struct VP9Common *cm);
 
+void vp9_swap_current_and_last_seg_map(struct VP9Common *cm);
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif

vp9/common/vp9_entropymode.c

@@ -439,9 +439,13 @@ void vp9_setup_past_independence(VP9_COMMON *cm) {
   int i;
   vp9_clearall_segfeatures(&cm->seg);
   cm->seg.abs_delta = SEGMENT_DELTADATA;
-  if (cm->last_frame_seg_map)
+
+  if (cm->last_frame_seg_map && !cm->frame_parallel_decode)
     vpx_memset(cm->last_frame_seg_map, 0, (cm->mi_rows * cm->mi_cols));
 
+  if (cm->current_frame_seg_map)
+    vpx_memset(cm->current_frame_seg_map, 0, (cm->mi_rows * cm->mi_cols));
+
   // Reset the mode ref deltas for loop filter
   vp9_zero(lf->last_ref_deltas);
   vp9_zero(lf->last_mode_deltas);
@@ -464,7 +468,7 @@ void vp9_setup_past_independence(VP9_COMMON *cm) {
     cm->frame_contexts[cm->frame_context_idx] = cm->fc;
   }
 
-  if (frame_is_intra_only(cm))
+  if (frame_is_intra_only(cm) && !cm->frame_parallel_decode)
     vpx_memset(cm->prev_mip, 0, cm->mi_stride * (cm->mi_rows + 1) *
                                     sizeof(*cm->prev_mip));
 

vp9/common/vp9_mvref_common.c

@@ -17,14 +17,12 @@ static void find_mv_refs_idx(const VP9_COMMON *cm, const MACROBLOCKD *xd,
                              const TileInfo *const tile,
                              MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
                              int_mv *mv_ref_list,
-                             int block, int mi_row, int mi_col) {
+                             int block, int mi_row, int mi_col,
+                             find_mv_refs_sync sync, void *const data) {
   const int *ref_sign_bias = cm->ref_frame_sign_bias;
   int i, refmv_count = 0;
-  const MODE_INFO *prev_mi = cm->coding_use_prev_mi && cm->prev_mi
-        ? cm->prev_mi_grid_visible[mi_row * xd->mi_stride + mi_col]
-        : NULL;
-  const MB_MODE_INFO *const prev_mbmi = prev_mi ? &prev_mi->mbmi : NULL;
-
+  MODE_INFO *prev_mi = NULL;
+  MB_MODE_INFO *prev_mbmi = NULL;
 
   const POSITION *const mv_ref_search = mv_ref_blocks[mi->mbmi.sb_type];
 
@@ -71,6 +69,14 @@ static void find_mv_refs_idx(const VP9_COMMON *cm, const MACROBLOCKD *xd,
     }
   }
 
+  // Synchronize here for frame parallel decode if sync function is provided.
+  if (sync != NULL) {
+    sync(data, mi_row);
+  }
+  prev_mi = cm->coding_use_prev_mi && cm->prev_mi ?
+            cm->prev_mi_grid_visible[mi_row * xd->mi_stride + mi_col] : NULL;
+  prev_mbmi = prev_mi ? &prev_mi->mbmi : NULL;
+
   // Check the last frame's mode and mv info.
   if (prev_mbmi) {
     if (prev_mbmi->ref_frame[0] == ref_frame)
@@ -112,9 +118,10 @@ void vp9_find_mv_refs(const VP9_COMMON *cm, const MACROBLOCKD *xd,
                       const TileInfo *const tile,
                       MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
                       int_mv *mv_ref_list,
-                                    int mi_row, int mi_col) {
+                      int mi_row, int mi_col,
+                      find_mv_refs_sync sync, void *const data) {
   find_mv_refs_idx(cm, xd, tile, mi, ref_frame, mv_ref_list, -1,
-                   mi_row, mi_col);
+                   mi_row, mi_col, sync, data);
 }
 
 static void lower_mv_precision(MV *mv, int allow_hp) {
@@ -152,7 +159,7 @@ void vp9_append_sub8x8_mvs_for_idx(VP9_COMMON *cm, MACROBLOCKD *xd,
   assert(MAX_MV_REF_CANDIDATES == 2);
 
   find_mv_refs_idx(cm, xd, tile, mi, mi->mbmi.ref_frame[ref], mv_list, block,
-                   mi_row, mi_col);
+                   mi_row, mi_col, NULL, NULL);
 
   near->as_int = 0;
   switch (block) {

vp9/common/vp9_mvref_common.h

@@ -204,10 +204,12 @@ static INLINE void clamp_mv2(MV *mv, const MACROBLOCKD *xd) {
                xd->mb_to_bottom_edge + RIGHT_BOTTOM_MARGIN);
 }
 
+typedef void (*find_mv_refs_sync)(void *const data, int mi_row);
 void vp9_find_mv_refs(const VP9_COMMON *cm, const MACROBLOCKD *xd,
                       const TileInfo *const tile,
                       MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
-                      int_mv *mv_ref_list, int mi_row, int mi_col);
+                      int_mv *mv_ref_list, int mi_row, int mi_col,
+                      find_mv_refs_sync sync, void *const data);
 
 // check a list of motion vectors by sad score using a number rows of pixels
 // above and a number cols of pixels in the left to select the one with best

vp9/common/vp9_onyxc_int.h

@@ -20,6 +20,7 @@
 #include "vp9/common/vp9_entropymode.h"
 #include "vp9/common/vp9_frame_buffers.h"
 #include "vp9/common/vp9_quant_common.h"
+#include "vp9/common/vp9_thread.h"
 #include "vp9/common/vp9_tile_common.h"
 
 #if CONFIG_VP9_POSTPROC
@@ -35,14 +36,19 @@ extern "C" {
 #define REF_FRAMES_LOG2 3
 #define REF_FRAMES (1 << REF_FRAMES_LOG2)
 
-// 1 scratch frame for the new frame, 3 for scaled references on the encoder
+// 4 scratch frames for the new frames to support a maximum of 4 cores decoding
+// in parallel, 3 for scaled references on the encoder.
+// TODO(hkuang): Add ondemand frame buffers instead of hardcoding the number
+// of framebuffers.
 // TODO(jkoleszar): These 3 extra references could probably come from the
 // normal reference pool.
-#define FRAME_BUFFERS (REF_FRAMES + 4)
+#define FRAME_BUFFERS (REF_FRAMES + 7)
 
 #define FRAME_CONTEXTS_LOG2 2
 #define FRAME_CONTEXTS (1 << FRAME_CONTEXTS_LOG2)
 
+#define NUM_PING_PONG_BUFFERS 2
+
 extern const struct {
   PARTITION_CONTEXT above;
   PARTITION_CONTEXT left;
@@ -61,8 +67,40 @@ typedef struct {
   int ref_count;
   vpx_codec_frame_buffer_t raw_frame_buffer;
   YV12_BUFFER_CONFIG buf;
+
+  // The Following variables will only be used in frame parallel decode.
+
+  // frame_worker_owner indicates which FrameWorker owns this buffer. NULL means
+  // that no FrameWorker owns, or is decoding, this buffer.
+  VP9Worker *frame_worker_owner;
+
+  // row and col indicate which position frame has been decoded to in real
+  // pixel unit. They are reset to -1 when decoding begins and set to INT_MAX
+  // when the frame is fully decoded.
+  int row;
+  int col;
 } RefCntBuffer;
 
+typedef struct {
+  // Protect BufferPool from being accessed by several FrameWorkers at
+  // the same time during frame parallel decode.
+  // TODO(hkuang): Try to use atomic variable instead of locking the whole pool.
+#if CONFIG_MULTITHREAD
+  pthread_mutex_t pool_mutex;
+#endif
+
+  // Private data associated with the frame buffer callbacks.
+  void *cb_priv;
+
+  vpx_get_frame_buffer_cb_fn_t get_fb_cb;
+  vpx_release_frame_buffer_cb_fn_t release_fb_cb;
+
+  RefCntBuffer frame_bufs[FRAME_BUFFERS];
+
+  // Frame buffers allocated internally by the codec.
+  InternalFrameBufferList int_frame_buffers;
+} BufferPool;
+
 typedef struct VP9Common {
   struct vpx_internal_error_info  error;
 
@@ -89,10 +127,12 @@ typedef struct VP9Common {
 
   YV12_BUFFER_CONFIG *frame_to_show;
 
-  RefCntBuffer frame_bufs[FRAME_BUFFERS];
-
   int ref_frame_map[REF_FRAMES]; /* maps fb_idx to reference slot */
 
+  // Prepare ref_frame_map for the next frame.
+  // Only used in frame parallel decode.
+  int next_ref_frame_map[REF_FRAMES];
+
   // TODO(jkoleszar): could expand active_ref_idx to 4, with 0 as intra, and
   // roll new_fb_idx into it.
 
@@ -144,8 +184,8 @@ typedef struct VP9Common {
 
   int mi_idx;
   int prev_mi_idx;
-  MODE_INFO *mip_array[2];
-  MODE_INFO **mi_grid_base_array[2];
+  MODE_INFO *mip_array[NUM_PING_PONG_BUFFERS];
+  MODE_INFO **mi_grid_base_array[NUM_PING_PONG_BUFFERS];
 
   MODE_INFO *mip; /* Base of allocated array */
   MODE_INFO *mi;  /* Corresponds to upper left visible macroblock */
@@ -157,8 +197,16 @@ typedef struct VP9Common {
   MODE_INFO **prev_mi_grid_base;
   MODE_INFO **prev_mi_grid_visible;
 
+  // Used in frame parallel decode for delay resizing prev_mi.
+  int update_prev_mi;
+
   // Persistent mb segment id map used in prediction.
-  unsigned char *last_frame_seg_map;
+  int seg_map_idx;
+  int prev_seg_map_idx;
+
+  uint8_t *seg_map_array[NUM_PING_PONG_BUFFERS];
+  uint8_t *last_frame_seg_map;
+  uint8_t *current_frame_seg_map;
 
   INTERP_FILTER interp_filter;
 
@@ -171,6 +219,10 @@ typedef struct VP9Common {
   struct loopfilter lf;
   struct segmentation seg;
 
+  // TODO(hkuang): Remove this as it is the same as frame_parallel_decode
+  // in pbi.
+  int frame_parallel_decode;  // frame-based threading.
+
   // Context probabilities for reference frame prediction
   int allow_comp_inter_inter;
   MV_REFERENCE_FRAME comp_fixed_ref;
@@ -202,30 +254,34 @@ typedef struct VP9Common {
 
   int log2_tile_cols, log2_tile_rows;
 
-  // Private data associated with the frame buffer callbacks.
-  void *cb_priv;
-  vpx_get_frame_buffer_cb_fn_t get_fb_cb;
-  vpx_release_frame_buffer_cb_fn_t release_fb_cb;
-
-  // Handles memory for the codec.
-  InternalFrameBufferList int_frame_buffers;
+  // External BufferPool passed from outside.
+  BufferPool *buffer_pool;
 
   PARTITION_CONTEXT *above_seg_context;
   ENTROPY_CONTEXT *above_context;
 } VP9_COMMON;
 
+// TODO(hkuang): Don't need to lock the whole pool after implementing atomic
+// frame reference count.
+void lock_buffer_pool(BufferPool *const pool);
+void unlock_buffer_pool(BufferPool *const pool);
+
 static INLINE YV12_BUFFER_CONFIG *get_frame_new_buffer(VP9_COMMON *cm) {
-  return &cm->frame_bufs[cm->new_fb_idx].buf;
+  return &cm->buffer_pool->frame_bufs[cm->new_fb_idx].buf;
 }
 
 static INLINE int get_free_fb(VP9_COMMON *cm) {
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
   int i;
-  for (i = 0; i < FRAME_BUFFERS; i++)
-    if (cm->frame_bufs[i].ref_count == 0)
+
+  lock_buffer_pool(cm->buffer_pool);
+  for (i = 0; i < FRAME_BUFFERS; ++i)
+    if (frame_bufs[i].ref_count == 0)
       break;
 
   assert(i < FRAME_BUFFERS);
-  cm->frame_bufs[i].ref_count = 1;
+  frame_bufs[i].ref_count = 1;
+  unlock_buffer_pool(cm->buffer_pool);
   return i;
 }
 
@@ -310,7 +366,6 @@ static INLINE void update_partition_context(MACROBLOCKD *xd,
   PARTITION_CONTEXT *const above_ctx = xd->above_seg_context + mi_col;
   PARTITION_CONTEXT *const left_ctx = xd->left_seg_context + (mi_row & MI_MASK);
 
-  // num_4x4_blocks_wide_lookup[bsize] / 2
   const int bs = num_8x8_blocks_wide_lookup[bsize];
 
   // update the partition context at the end notes. set partition bits

vp9/common/vp9_reconinter.c

@@ -20,50 +20,7 @@
 #include "vp9/common/vp9_reconinter.h"
 #include "vp9/common/vp9_reconintra.h"
 
-static void build_mc_border(const uint8_t *src, int src_stride,
-                            uint8_t *dst, int dst_stride,
-                            int x, int y, int b_w, int b_h, int w, int h) {
-  // Get a pointer to the start of the real data for this row.
-  const uint8_t *ref_row = src - x - y * src_stride;
-
-  if (y >= h)
-    ref_row += (h - 1) * src_stride;
-  else if (y > 0)
-    ref_row += y * src_stride;
-
-  do {
-    int right = 0, copy;
-    int left = x < 0 ? -x : 0;
-
-    if (left > b_w)
-      left = b_w;
-
-    if (x + b_w > w)
-      right = x + b_w - w;
-
-    if (right > b_w)
-      right = b_w;
-
-    copy = b_w - left - right;
-
-    if (left)
-      memset(dst, ref_row[0], left);
-
-    if (copy)
-      memcpy(dst + left, ref_row + x + left, copy);
-
-    if (right)
-      memset(dst + left + copy, ref_row[w - 1], right);
-
-    dst += dst_stride;
-    ++y;
-
-    if (y > 0 && y < h)
-      ref_row += src_stride;
-  } while (--b_h);
-}
-
-static void inter_predictor(const uint8_t *src, int src_stride,
+void inter_predictor(const uint8_t *src, int src_stride,
                             uint8_t *dst, int dst_stride,
                             const int subpel_x,
                             const int subpel_y,
@@ -151,7 +108,7 @@ MV clamp_mv_to_umv_border_sb(const MACROBLOCKD *xd, const MV *src_mv,
   return clamped_mv;
 }
 
-static MV average_split_mvs(const struct macroblockd_plane *pd, int plane,
+MV average_split_mvs(const struct macroblockd_plane *pd, int plane,
                             const MODE_INFO *mi, int ref, int block) {
   const int ss_idx = ((pd->subsampling_x > 0) << 1) | (pd->subsampling_y > 0);
   MV res = {0, 0};
@@ -174,7 +131,7 @@ static MV average_split_mvs(const struct macroblockd_plane *pd, int plane,
   return res;
 }
 
-static void build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
+void build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
                                    int bw, int bh,
                                    int x, int y, int w, int h,
                                    int mi_x, int mi_y) {
@@ -270,169 +227,6 @@ void vp9_build_inter_predictors_sb(MACROBLOCKD *xd, int mi_row, int mi_col,
                                     MAX_MB_PLANE - 1);
 }
 
-// TODO(jingning): This function serves as a placeholder for decoder prediction
-// using on demand border extension. It should be moved to /decoder/ directory.
-static void dec_build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
-                                       int bw, int bh,
-                                       int x, int y, int w, int h,
-                                       int mi_x, int mi_y) {
-  struct macroblockd_plane *const pd = &xd->plane[plane];
-  const MODE_INFO *mi = xd->mi[0];
-  const int is_compound = has_second_ref(&mi->mbmi);
-  const InterpKernel *kernel = vp9_get_interp_kernel(mi->mbmi.interp_filter);
-  int ref;
-
-  for (ref = 0; ref < 1 + is_compound; ++ref) {
-    const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
-    struct buf_2d *const pre_buf = &pd->pre[ref];
-    struct buf_2d *const dst_buf = &pd->dst;
-    uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
-    const MV mv = mi->mbmi.sb_type < BLOCK_8X8
-               ? average_split_mvs(pd, plane, mi, ref, block)
-               : mi->mbmi.mv[ref].as_mv;
-
-
-    // TODO(jkoleszar): This clamping is done in the incorrect place for the
-    // scaling case. It needs to be done on the scaled MV, not the pre-scaling
-    // MV. Note however that it performs the subsampling aware scaling so
-    // that the result is always q4.
-    // mv_precision precision is MV_PRECISION_Q4.
-    const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, &mv, bw, bh,
-                                               pd->subsampling_x,
-                                               pd->subsampling_y);
-
-    MV32 scaled_mv;
-    int xs, ys, x0, y0, x0_16, y0_16, frame_width, frame_height, buf_stride,
-        subpel_x, subpel_y;
-    uint8_t *ref_frame, *buf_ptr;
-    const YV12_BUFFER_CONFIG *ref_buf = xd->block_refs[ref]->buf;
-
-    // Get reference frame pointer, width and height.
-    if (plane == 0) {
-      frame_width = ref_buf->y_crop_width;
-      frame_height = ref_buf->y_crop_height;
-      ref_frame = ref_buf->y_buffer;
-    } else {
-      frame_width = ref_buf->uv_crop_width;
-      frame_height = ref_buf->uv_crop_height;
-      ref_frame = plane == 1 ? ref_buf->u_buffer : ref_buf->v_buffer;
-    }
-
-    if (vp9_is_scaled(sf)) {
-      // Co-ordinate of containing block to pixel precision.
-      int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
-      int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
-
-      // Co-ordinate of the block to 1/16th pixel precision.
-      x0_16 = (x_start + x) << SUBPEL_BITS;
-      y0_16 = (y_start + y) << SUBPEL_BITS;
-
-      // Co-ordinate of current block in reference frame
-      // to 1/16th pixel precision.
-      x0_16 = sf->scale_value_x(x0_16, sf);
-      y0_16 = sf->scale_value_y(y0_16, sf);
-
-      // Map the top left corner of the block into the reference frame.
-      x0 = sf->scale_value_x(x_start + x, sf);
-      y0 = sf->scale_value_y(y_start + y, sf);
-
-      // Scale the MV and incorporate the sub-pixel offset of the block
-      // in the reference frame.
-      scaled_mv = vp9_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
-      xs = sf->x_step_q4;
-      ys = sf->y_step_q4;
-    } else {
-      // Co-ordinate of containing block to pixel precision.
-      x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
-      y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
-
-      // Co-ordinate of the block to 1/16th pixel precision.
-      x0_16 = x0 << SUBPEL_BITS;
-      y0_16 = y0 << SUBPEL_BITS;
-
-      scaled_mv.row = mv_q4.row;
-      scaled_mv.col = mv_q4.col;
-      xs = ys = 16;
-    }
-    subpel_x = scaled_mv.col & SUBPEL_MASK;
-    subpel_y = scaled_mv.row & SUBPEL_MASK;
-
-    // Calculate the top left corner of the best matching block in the reference frame.
-    x0 += scaled_mv.col >> SUBPEL_BITS;
-    y0 += scaled_mv.row >> SUBPEL_BITS;
-    x0_16 += scaled_mv.col;
-    y0_16 += scaled_mv.row;
-
-    // Get reference block pointer.
-    buf_ptr = ref_frame + y0 * pre_buf->stride + x0;
-    buf_stride = pre_buf->stride;
-
-    // Do border extension if there is motion or the
-    // width/height is not a multiple of 8 pixels.
-    if (scaled_mv.col || scaled_mv.row ||
-        (frame_width & 0x7) || (frame_height & 0x7)) {
-      // Get reference block bottom right coordinate.
-      int x1 = ((x0_16 + (w - 1) * xs) >> SUBPEL_BITS) + 1;
-      int y1 = ((y0_16 + (h - 1) * ys) >> SUBPEL_BITS) + 1;
-      int x_pad = 0, y_pad = 0;
-
-      if (subpel_x || (sf->x_step_q4 & SUBPEL_MASK)) {
-        x0 -= VP9_INTERP_EXTEND - 1;
-        x1 += VP9_INTERP_EXTEND;
-        x_pad = 1;
-      }
-
-      if (subpel_y || (sf->y_step_q4 & SUBPEL_MASK)) {
-        y0 -= VP9_INTERP_EXTEND - 1;
-        y1 += VP9_INTERP_EXTEND;
-        y_pad = 1;
-      }
-
-      // Skip border extension if block is inside the frame.
-      if (x0 < 0 || x0 > frame_width - 1 || x1 < 0 || x1 > frame_width ||
-          y0 < 0 || y0 > frame_height - 1 || y1 < 0 || y1 > frame_height - 1) {
-        uint8_t *buf_ptr1 = ref_frame + y0 * pre_buf->stride + x0;
-        // Extend the border.
-        build_mc_border(buf_ptr1, pre_buf->stride, xd->mc_buf, x1 - x0 + 1,
-                        x0, y0, x1 - x0 + 1, y1 - y0 + 1, frame_width,
-                        frame_height);
-        buf_stride = x1 - x0 + 1;
-        buf_ptr = xd->mc_buf + y_pad * 3 * buf_stride + x_pad * 3;
-      }
-    }
-
-    inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
-                    subpel_y, sf, w, h, ref, kernel, xs, ys);
-  }
-}
-
-void vp9_dec_build_inter_predictors_sb(MACROBLOCKD *xd, int mi_row, int mi_col,
-                                       BLOCK_SIZE bsize) {
-  int plane;
-  const int mi_x = mi_col * MI_SIZE;
-  const int mi_y = mi_row * MI_SIZE;
-  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
-    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize,
-                                                        &xd->plane[plane]);
-    const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
-    const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
-    const int bw = 4 * num_4x4_w;
-    const int bh = 4 * num_4x4_h;
-
-    if (xd->mi[0]->mbmi.sb_type < BLOCK_8X8) {
-      int i = 0, x, y;
-      assert(bsize == BLOCK_8X8);
-      for (y = 0; y < num_4x4_h; ++y)
-        for (x = 0; x < num_4x4_w; ++x)
-          dec_build_inter_predictors(xd, plane, i++, bw, bh,
-                                     4 * x, 4 * y, 4, 4, mi_x, mi_y);
-    } else {
-      dec_build_inter_predictors(xd, plane, 0, bw, bh,
-                                 0, 0, bw, bh, mi_x, mi_y);
-    }
-  }
-}
-
 void vp9_setup_dst_planes(struct macroblockd_plane planes[MAX_MB_PLANE],
                           const YV12_BUFFER_CONFIG *src,
                           int mi_row, int mi_col) {

vp9/common/vp9_reconinter.h

@@ -18,6 +18,26 @@
 extern "C" {
 #endif
 
+void inter_predictor(const uint8_t *src, int src_stride,
+                            uint8_t *dst, int dst_stride,
+                            const int subpel_x,
+                            const int subpel_y,
+                            const struct scale_factors *sf,
+                            int w, int h, int ref,
+                            const InterpKernel *kernel,
+                            int xs, int ys);
+
+MV average_split_mvs(const struct macroblockd_plane *pd, int plane,
+                            const MODE_INFO *mi, int ref, int block);
+
+MV clamp_mv_to_umv_border_sb(const MACROBLOCKD *xd, const MV *src_mv,
+                             int bw, int bh, int ss_x, int ss_y);
+
+void build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
+                                   int bw, int bh,
+                                   int x, int y, int w, int h,
+                                   int mi_x, int mi_y);
+
 void vp9_build_inter_predictors_sby(MACROBLOCKD *xd, int mi_row, int mi_col,
                                     BLOCK_SIZE bsize);
 
@@ -27,9 +47,6 @@ void vp9_build_inter_predictors_sbuv(MACROBLOCKD *xd, int mi_row, int mi_col,
 void vp9_build_inter_predictors_sb(MACROBLOCKD *xd, int mi_row, int mi_col,
                                    BLOCK_SIZE bsize);
 
-void vp9_dec_build_inter_predictors_sb(MACROBLOCKD *xd, int mi_row, int mi_col,
-                                       BLOCK_SIZE bsize);
-
 void vp9_build_inter_predictor(const uint8_t *src, int src_stride,
                                uint8_t *dst, int dst_stride,
                                const MV *mv_q3,

vp9/decoder/vp9_decodeframe.c

@@ -28,6 +28,7 @@
 #include "vp9/common/vp9_reconintra.h"
 #include "vp9/common/vp9_reconinter.h"
 #include "vp9/common/vp9_seg_common.h"
+#include "vp9/common/vp9_thread.h"
 #include "vp9/common/vp9_tile_common.h"
 
 #include "vp9/decoder/vp9_decodeframe.h"
@@ -38,7 +39,6 @@
 #include "vp9/decoder/vp9_dthread.h"
 #include "vp9/decoder/vp9_read_bit_buffer.h"
 #include "vp9/decoder/vp9_reader.h"
-#include "vp9/decoder/vp9_thread.h"
 
 #define MAX_VP9_HEADER_SIZE 80
 
@@ -327,21 +327,24 @@ static void set_ref(VP9_COMMON *const cm, MACROBLOCKD *const xd,
   MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
   RefBuffer *ref_buffer = &cm->frame_refs[mbmi->ref_frame[idx] - LAST_FRAME];
   xd->block_refs[idx] = ref_buffer;
+
   if (!vp9_is_valid_scale(&ref_buffer->sf))
     vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                        "Invalid scale factors");
   vp9_setup_pre_planes(xd, idx, ref_buffer->buf, mi_row, mi_col,
                        &ref_buffer->sf);
+  if (!cm->frame_parallel_decode)
     xd->corrupted |= ref_buffer->buf->corrupted;
 }
 
-static void decode_block(VP9_COMMON *const cm, MACROBLOCKD *const xd,
+static void decode_block(VP9Decoder *const pbi, MACROBLOCKD *const xd,
                          const TileInfo *const tile,
                          int mi_row, int mi_col,
                          vp9_reader *r, BLOCK_SIZE bsize) {
+  VP9_COMMON *const cm = &pbi->common;
   const int less8x8 = bsize < BLOCK_8X8;
   MB_MODE_INFO *mbmi = set_offsets(cm, xd, tile, bsize, mi_row, mi_col);
-  vp9_read_mode_info(cm, xd, tile, mi_row, mi_col, r);
+  vp9_read_mode_info(pbi, xd, tile, mi_row, mi_col, r);
 
   if (less8x8)
     bsize = BLOCK_8X8;
@@ -365,7 +368,7 @@ static void decode_block(VP9_COMMON *const cm, MACROBLOCKD *const xd,
       set_ref(cm, xd, 1, mi_row, mi_col);
 
     // Prediction
-    vp9_dec_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
+    vp9_dec_build_inter_predictors_sb(pbi, xd, mi_row, mi_col, bsize);
 
     // Reconstruction
     if (!mbmi->skip) {
@@ -404,10 +407,11 @@ static PARTITION_TYPE read_partition(VP9_COMMON *cm, MACROBLOCKD *xd, int hbs,
   return p;
 }
 
-static void decode_partition(VP9_COMMON *const cm, MACROBLOCKD *const xd,
+static void decode_partition(VP9Decoder *const pbi, MACROBLOCKD *const xd,
                              const TileInfo *const tile,
                              int mi_row, int mi_col,
                              vp9_reader* r, BLOCK_SIZE bsize) {
+  VP9_COMMON *const cm = &pbi->common;
   const int hbs = num_8x8_blocks_wide_lookup[bsize] / 2;
   PARTITION_TYPE partition;
   BLOCK_SIZE subsize, uv_subsize;
@@ -422,27 +426,27 @@ static void decode_partition(VP9_COMMON *const cm, MACROBLOCKD *const xd,
     vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
                        "Invalid block size.");
   if (subsize < BLOCK_8X8) {
-    decode_block(cm, xd, tile, mi_row, mi_col, r, subsize);
+    decode_block(pbi, xd, tile, mi_row, mi_col, r, subsize);
   } else {
     switch (partition) {
       case PARTITION_NONE:
-        decode_block(cm, xd, tile, mi_row, mi_col, r, subsize);
+        decode_block(pbi, xd, tile, mi_row, mi_col, r, subsize);
         break;
       case PARTITION_HORZ:
-        decode_block(cm, xd, tile, mi_row, mi_col, r, subsize);
+        decode_block(pbi, xd, tile, mi_row, mi_col, r, subsize);
         if (mi_row + hbs < cm->mi_rows)
-          decode_block(cm, xd, tile, mi_row + hbs, mi_col, r, subsize);
+          decode_block(pbi, xd, tile, mi_row + hbs, mi_col, r, subsize);
         break;
       case PARTITION_VERT:
-        decode_block(cm, xd, tile, mi_row, mi_col, r, subsize);
+        decode_block(pbi, xd, tile, mi_row, mi_col, r, subsize);
         if (mi_col + hbs < cm->mi_cols)
-          decode_block(cm, xd, tile, mi_row, mi_col + hbs, r, subsize);
+          decode_block(pbi, xd, tile, mi_row, mi_col + hbs, r, subsize);
         break;
       case PARTITION_SPLIT:
-        decode_partition(cm, xd, tile, mi_row,       mi_col,       r, subsize);
-        decode_partition(cm, xd, tile, mi_row,       mi_col + hbs, r, subsize);
-        decode_partition(cm, xd, tile, mi_row + hbs, mi_col,       r, subsize);
-        decode_partition(cm, xd, tile, mi_row + hbs, mi_col + hbs, r, subsize);
+        decode_partition(pbi, xd, tile, mi_row,       mi_col,       r, subsize);
+        decode_partition(pbi, xd, tile, mi_row,       mi_col + hbs, r, subsize);
+        decode_partition(pbi, xd, tile, mi_row + hbs, mi_col,       r, subsize);
+        decode_partition(pbi, xd, tile, mi_row + hbs, mi_col + hbs, r, subsize);
         break;
       default:
         assert(0 && "Invalid partition type");
@@ -621,6 +625,7 @@ static void setup_display_size(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
 }
 
 static void apply_frame_size(VP9_COMMON *cm, int width, int height) {
+  BufferPool *const pool = cm->buffer_pool;
   if (cm->width != width || cm->height != height) {
     // Change in frame size.
     // TODO(agrange) Don't test width/height, check overall size.
@@ -637,14 +642,17 @@ static void apply_frame_size(VP9_COMMON *cm, int width, int height) {
     vp9_update_frame_size(cm);
   }
 
+  lock_buffer_pool(pool);
   if (vp9_realloc_frame_buffer(
           get_frame_new_buffer(cm), cm->width, cm->height,
           cm->subsampling_x, cm->subsampling_y, VP9_DEC_BORDER_IN_PIXELS,
-          &cm->frame_bufs[cm->new_fb_idx].raw_frame_buffer, cm->get_fb_cb,
-          cm->cb_priv)) {
+          &pool->frame_bufs[cm->new_fb_idx].raw_frame_buffer, pool->get_fb_cb,
+          pool->cb_priv)) {
+    unlock_buffer_pool(pool);
     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                        "Failed to allocate frame buffer");
   }
+  unlock_buffer_pool(pool);
 }
 
 static void setup_frame_size(VP9_COMMON *cm, struct vp9_read_bit_buffer *rb) {
@@ -777,7 +785,7 @@ static const uint8_t *decode_tiles(VP9Decoder *pbi,
   const int tile_rows = 1 << cm->log2_tile_rows;
   TileBuffer tile_buffers[4][1 << 6];
   int tile_row, tile_col;
-  int mi_row, mi_col;
+  int mi_row = 0, mi_col = 0;
   TileData *tile_data = NULL;
 
   if (cm->lf.filter_level && pbi->lf_worker.data1 == NULL) {
@@ -797,7 +805,6 @@ static const uint8_t *decode_tiles(VP9Decoder *pbi,
     vp9_copy(lf_data->planes, pbi->mb.plane);
     lf_data->stop = 0;
     lf_data->y_only = 0;
-    vp9_loop_filter_frame_init(cm, cm->lf.filter_level);
   }
 
   assert(tile_rows <= 4);
@@ -855,7 +862,7 @@ static const uint8_t *decode_tiles(VP9Decoder *pbi,
         vp9_zero(tile_data->xd.left_seg_context);
         for (mi_col = tile.mi_col_start; mi_col < tile.mi_col_end;
              mi_col += MI_BLOCK_SIZE) {
-          decode_partition(tile_data->cm, &tile_data->xd, &tile, mi_row, mi_col,
+          decode_partition(pbi, &tile_data->xd, &tile, mi_row, mi_col,
                            &tile_data->bit_reader, BLOCK_64X64);
         }
       }
@@ -879,6 +886,12 @@ static const uint8_t *decode_tiles(VP9Decoder *pbi,
           winterface->execute(&pbi->lf_worker);
         }
       }
+      // After loopfiltering, the last 7 row pixels in each superblock row may
+      // still be changed by the longest loopfilter of the next superblock
+      // row.
+      if (pbi->frame_parallel_decode)
+        vp9_frameworker_broadcast(pbi->cur_buf,
+                                  mi_row << MI_BLOCK_SIZE_LOG2);
     }
   }
 
@@ -894,6 +907,8 @@ static const uint8_t *decode_tiles(VP9Decoder *pbi,
   // Get last tile data.
   tile_data = pbi->tile_data + tile_cols * tile_rows - 1;
 
+  if (pbi->frame_parallel_decode)
+    vp9_frameworker_broadcast(pbi->cur_buf, INT_MAX);
   return vp9_reader_find_end(&tile_data->bit_reader);
 }
 
@@ -908,7 +923,7 @@ static int tile_worker_hook(void *arg1, void *arg2) {
     vp9_zero(tile_data->xd.left_seg_context);
     for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
          mi_col += MI_BLOCK_SIZE) {
-      decode_partition(tile_data->cm, &tile_data->xd, tile,
+      decode_partition(tile_data->pbi, &tile_data->xd, tile,
                        mi_row, mi_col, &tile_data->bit_reader, BLOCK_64X64);
     }
   }
@@ -1014,10 +1029,10 @@ static const uint8_t *decode_tiles_mt(VP9Decoder *pbi,
       TileInfo *const tile = (TileInfo*)worker->data2;
       TileBuffer *const buf = &tile_buffers[0][n];
 
-      tile_data->cm = cm;
+      tile_data->pbi = pbi;
       tile_data->xd = pbi->mb;
       tile_data->xd.corrupted = 0;
-      vp9_tile_init(tile, tile_data->cm, 0, buf->col);
+      vp9_tile_init(tile, &pbi->common, 0, buf->col);
       setup_token_decoder(buf->data, data_end, buf->size, &cm->error,
                           &tile_data->bit_reader, pbi->decrypt_cb,
                           pbi->decrypt_state);
@@ -1076,8 +1091,10 @@ static BITSTREAM_PROFILE read_profile(struct vp9_read_bit_buffer *rb) {
 static size_t read_uncompressed_header(VP9Decoder *pbi,
                                        struct vp9_read_bit_buffer *rb) {
   VP9_COMMON *const cm = &pbi->common;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+  BufferPool *const pool = pbi->common.buffer_pool;
+  int i, mask, ref_index = 0;
   size_t sz;
-  int i;
 
   cm->last_frame_type = cm->frame_type;
 
@@ -1094,16 +1111,22 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
   if (cm->show_existing_frame) {
     // Show an existing frame directly.
     const int frame_to_show = cm->ref_frame_map[vp9_rb_read_literal(rb, 3)];
-
-    if (frame_to_show < 0 || cm->frame_bufs[frame_to_show].ref_count < 1)
+    lock_buffer_pool(pool);
+    if (frame_to_show < 0 || frame_bufs[frame_to_show].ref_count < 1)
       vpx_internal_error(&cm->error, VPX_CODEC_UNSUP_BITSTREAM,
                          "Buffer %d does not contain a decoded frame",
                          frame_to_show);
 
-    ref_cnt_fb(cm->frame_bufs, &cm->new_fb_idx, frame_to_show);
+    ref_cnt_fb(frame_bufs, &cm->new_fb_idx, frame_to_show);
+    unlock_buffer_pool(pool);
     pbi->refresh_frame_flags = 0;
     cm->lf.filter_level = 0;
     cm->show_frame = 1;
+
+    if (pbi->frame_parallel_decode) {
+      for (i = 0; i < REF_FRAMES; ++i)
+        cm->next_ref_frame_map[i] = cm->ref_frame_map[i];
+    }
     return 0;
   }
 
@@ -1143,6 +1166,10 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
     }
 
     setup_frame_size(cm, rb);
+    if (pbi->need_resync) {
+      vpx_memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+      pbi->need_resync = 0;
+    }
   } else {
     cm->intra_only = cm->show_frame ? 0 : vp9_rb_read_bit(rb);
 
@@ -1154,6 +1181,10 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
 
       pbi->refresh_frame_flags = vp9_rb_read_literal(rb, REF_FRAMES);
       setup_frame_size(cm, rb);
+      if (pbi->need_resync) {
+        vpx_memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+        pbi->need_resync = 0;
+      }
     } else {
       pbi->refresh_frame_flags = vp9_rb_read_literal(rb, REF_FRAMES);
       for (i = 0; i < REFS_PER_FRAME; ++i) {
@@ -1161,10 +1192,9 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
         const int idx = cm->ref_frame_map[ref];
         RefBuffer *const ref_frame = &cm->frame_refs[i];
         ref_frame->idx = idx;
-        ref_frame->buf = &cm->frame_bufs[idx].buf;
+        ref_frame->buf = &frame_bufs[idx].buf;
         cm->ref_frame_sign_bias[LAST_FRAME + i] = vp9_rb_read_bit(rb);
       }
-
       setup_frame_size_with_refs(cm, rb);
 
       cm->allow_high_precision_mv = vp9_rb_read_bit(rb);
@@ -1182,6 +1212,12 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
     }
   }
 
+  if (pbi->need_resync) {
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Keyframe / intra-only frame required to reset decoder"
+                       " state");
+  }
+
   if (!cm->error_resilient_mode) {
     cm->coding_use_prev_mi = 1;
     cm->refresh_frame_context = vp9_rb_read_bit(rb);
@@ -1196,6 +1232,30 @@ static size_t read_uncompressed_header(VP9Decoder *pbi,
   // below, forcing the use of context 0 for those frame types.
   cm->frame_context_idx = vp9_rb_read_literal(rb, FRAME_CONTEXTS_LOG2);
 
+  // Generate next_ref_frame_map.
+  lock_buffer_pool(pool);
+  for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
+    if (mask & 1) {
+      cm->next_ref_frame_map[ref_index] = cm->new_fb_idx;
+      ++frame_bufs[cm->new_fb_idx].ref_count;
+    } else {
+      cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
+    }
+    // Current thread holds the reference frame.
+    if (cm->ref_frame_map[ref_index] >= 0)
+      ++frame_bufs[cm->ref_frame_map[ref_index]].ref_count;
+    ++ref_index;
+  }
+
+  for (; ref_index < REF_FRAMES; ++ref_index) {
+    cm->next_ref_frame_map[ref_index] = cm->ref_frame_map[ref_index];
+    // Current thread holds the reference frame.
+    if (cm->ref_frame_map[ref_index] >= 0)
+      ++frame_bufs[cm->ref_frame_map[ref_index]].ref_count;
+  }
+  unlock_buffer_pool(pool);
+  pbi->hold_ref_buf = 1;
+
   if (frame_is_intra_only(cm) || cm->error_resilient_mode)
     vp9_setup_past_independence(cm);
 
@@ -1341,6 +1401,7 @@ void vp9_decode_frame(VP9Decoder *pbi,
   VP9_COMMON *const cm = &pbi->common;
   MACROBLOCKD *const xd = &pbi->mb;
   struct vp9_read_bit_buffer rb = { NULL, NULL, 0, NULL, 0};
+  int context_updated = 0;
 
   uint8_t clear_data[MAX_VP9_HEADER_SIZE];
   const size_t first_partition_size = read_uncompressed_header(pbi,
@@ -1378,6 +1439,28 @@ void vp9_decode_frame(VP9Decoder *pbi,
   xd->corrupted = 0;
   new_fb->corrupted = read_compressed_header(pbi, data, first_partition_size);
 
+  if (cm->lf.filter_level) {
+    vp9_loop_filter_frame_init(cm, cm->lf.filter_level);
+  }
+
+  // If encoded in frame parallel mode, frame context is ready after decoding
+  // the frame header.
+  if (pbi->frame_parallel_decode && cm->frame_parallel_decoding_mode) {
+    VP9Worker *const worker = pbi->frame_worker_owner;
+    FrameWorkerData *const frame_worker_data = worker->data1;
+    if (cm->refresh_frame_context) {
+      context_updated = 1;
+      cm->frame_contexts[cm->frame_context_idx] = cm->fc;
+    }
+    vp9_frameworker_lock_stats(worker);
+    pbi->cur_buf->row = -1;
+    pbi->cur_buf->col = -1;
+    frame_worker_data->frame_context_ready = 1;
+    // Signal the main thread that context is ready.
+    vp9_frameworker_signal_stats(worker);
+    vp9_frameworker_unlock_stats(worker);
+  }
+
   // TODO(jzern): remove frame_parallel_decoding_mode restriction for
   // single-frame tile decoding.
   if (pbi->max_threads > 1 && tile_rows == 1 && tile_cols > 1 &&
@@ -1390,9 +1473,7 @@ void vp9_decode_frame(VP9Decoder *pbi,
     *p_data_end = decode_tiles(pbi, data + first_partition_size, data_end);
   }
 
-  new_fb->corrupted |= xd->corrupted;
-
-  if (!new_fb->corrupted) {
+  if (!xd->corrupted) {
     if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
       vp9_adapt_coef_probs(cm);
 
@@ -1403,8 +1484,235 @@ void vp9_decode_frame(VP9Decoder *pbi,
     } else {
       debug_check_frame_counts(cm);
     }
+  } else {
+    vpx_internal_error(&cm->error, VPX_CODEC_CORRUPT_FRAME,
+                       "Decode failed. Frame data is corrupted.");
   }
 
-  if (cm->refresh_frame_context)
+  // Non frame parallel update frame context here.
+  if (cm->refresh_frame_context && !context_updated)
     cm->frame_contexts[cm->frame_context_idx] = cm->fc;
 }
+
+static void build_mc_border(const uint8_t *src, int src_stride,
+                            uint8_t *dst, int dst_stride,
+                            int x, int y, int b_w, int b_h, int w, int h) {
+  // Get a pointer to the start of the real data for this row.
+  const uint8_t *ref_row = src - x - y * src_stride;
+
+  if (y >= h)
+    ref_row += (h - 1) * src_stride;
+  else if (y > 0)
+    ref_row += y * src_stride;
+
+  do {
+    int right = 0, copy;
+    int left = x < 0 ? -x : 0;
+
+    if (left > b_w)
+      left = b_w;
+
+    if (x + b_w > w)
+      right = x + b_w - w;
+
+    if (right > b_w)
+      right = b_w;
+
+    copy = b_w - left - right;
+
+    if (left)
+      memset(dst, ref_row[0], left);
+
+    if (copy)
+      memcpy(dst + left, ref_row + x + left, copy);
+
+    if (right)
+      memset(dst + left + copy, ref_row[w - 1], right);
+
+    dst += dst_stride;
+    ++y;
+
+    if (y > 0 && y < h)
+      ref_row += src_stride;
+  } while (--b_h);
+}
+
+void dec_build_inter_predictors(VP9Decoder *const pbi, MACROBLOCKD *xd,
+                                int plane, int block, int bw, int bh, int x,
+                                int y, int w, int h, int mi_x, int mi_y) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  const MODE_INFO *mi = xd->mi[0];
+  const int is_compound = has_second_ref(&mi->mbmi);
+  const InterpKernel *kernel = vp9_get_interp_kernel(mi->mbmi.interp_filter);
+  int ref;
+
+  for (ref = 0; ref < 1 + is_compound; ++ref) {
+    const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
+    struct buf_2d *const pre_buf = &pd->pre[ref];
+    struct buf_2d *const dst_buf = &pd->dst;
+    uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
+    const MV mv = mi->mbmi.sb_type < BLOCK_8X8
+               ? average_split_mvs(pd, plane, mi, ref, block)
+               : mi->mbmi.mv[ref].as_mv;
+
+
+    // TODO(jkoleszar): This clamping is done in the incorrect place for the
+    // scaling case. It needs to be done on the scaled MV, not the pre-scaling
+    // MV. Note however that it performs the subsampling aware scaling so
+    // that the result is always q4.
+    // mv_precision precision is MV_PRECISION_Q4.
+    const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, &mv, bw, bh,
+                                               pd->subsampling_x,
+                                               pd->subsampling_y);
+
+    MV32 scaled_mv;
+    int xs, ys, x0, y0, x0_16, y0_16, y1, frame_width, frame_height,
+        buf_stride, subpel_x, subpel_y;
+    uint8_t *ref_frame, *buf_ptr;
+    const int idx = xd->block_refs[ref]->idx;
+    BufferPool *const pool = pbi->common.buffer_pool;
+    RefCntBuffer *const ref_frame_buf = &pool->frame_bufs[idx];
+
+    // Get reference frame pointer, width and height.
+    if (plane == 0) {
+      frame_width = ref_frame_buf->buf.y_crop_width;
+      frame_height = ref_frame_buf->buf.y_crop_height;
+      ref_frame = ref_frame_buf->buf.y_buffer;
+    } else {
+      frame_width = ref_frame_buf->buf.uv_crop_width;
+      frame_height = ref_frame_buf->buf.uv_crop_height;
+      ref_frame = plane == 1 ? ref_frame_buf->buf.u_buffer
+                           : ref_frame_buf->buf.v_buffer;
+    }
+
+    if (vp9_is_scaled(sf)) {
+      // Co-ordinate of containing block to pixel precision.
+      int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
+      int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
+
+      // Co-ordinate of the block to 1/16th pixel precision.
+      x0_16 = (x_start + x) << SUBPEL_BITS;
+      y0_16 = (y_start + y) << SUBPEL_BITS;
+
+      // Co-ordinate of current block in reference frame
+      // to 1/16th pixel precision.
+      x0_16 = sf->scale_value_x(x0_16, sf);
+      y0_16 = sf->scale_value_y(y0_16, sf);
+
+      // Map the top left corner of the block into the reference frame.
+      x0 = sf->scale_value_x(x_start + x, sf);
+      y0 = sf->scale_value_y(y_start + y, sf);
+
+      // Scale the MV and incorporate the sub-pixel offset of the block
+      // in the reference frame.
+      scaled_mv = vp9_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
+      xs = sf->x_step_q4;
+      ys = sf->y_step_q4;
+    } else {
+      // Co-ordinate of containing block to pixel precision.
+      x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
+      y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
+
+      // Co-ordinate of the block to 1/16th pixel precision.
+      x0_16 = x0 << SUBPEL_BITS;
+      y0_16 = y0 << SUBPEL_BITS;
+
+      scaled_mv.row = mv_q4.row;
+      scaled_mv.col = mv_q4.col;
+      xs = ys = 16;
+    }
+    subpel_x = scaled_mv.col & SUBPEL_MASK;
+    subpel_y = scaled_mv.row & SUBPEL_MASK;
+
+    // Calculate the top left corner of the best matching block in the
+    // reference frame.
+    x0 += scaled_mv.col >> SUBPEL_BITS;
+    y0 += scaled_mv.row >> SUBPEL_BITS;
+    x0_16 += scaled_mv.col;
+    y0_16 += scaled_mv.row;
+
+    // Get reference block pointer.
+    buf_ptr = ref_frame + y0 * pre_buf->stride + x0;
+    buf_stride = pre_buf->stride;
+
+    // Get reference block bottom right vertical coordinate.
+    y1 = ((y0_16 + (h - 1) * ys) >> SUBPEL_BITS) + 1;
+
+    // Do border extension if there is motion or the
+    // width/height is not a multiple of 8 pixels.
+    if (scaled_mv.col || scaled_mv.row ||
+        (frame_width & 0x7) || (frame_height & 0x7)) {
+      int x_pad = 0, y_pad = 0;
+
+      // Get reference block bottom right horizontal coordinate.
+      int x1 = ((x0_16 + (w - 1) * xs) >> SUBPEL_BITS) + 1;
+
+      if (subpel_x || (sf->x_step_q4 & SUBPEL_MASK)) {
+        x0 -= VP9_INTERP_EXTEND - 1;
+        x1 += VP9_INTERP_EXTEND;
+        x_pad = 1;
+      }
+
+      if (subpel_y || (sf->y_step_q4 & SUBPEL_MASK)) {
+        y0 -= VP9_INTERP_EXTEND - 1;
+        y1 += VP9_INTERP_EXTEND;
+        y_pad = 1;
+      }
+
+      // Wait until reference block is ready. Pad 7 more pixels as last 7
+      // pixels of each superblock row can be changed by next superblock row.
+       if (pbi->frame_parallel_decode)
+         vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
+                              (y1 + 7) << (plane == 0 ? 0 : 1));
+
+      // Skip border extension if block is inside the frame.
+      if (x0 < 0 || x0 > frame_width - 1 || x1 < 0 || x1 > frame_width ||
+          y0 < 0 || y0 > frame_height - 1 || y1 < 0 || y1 > frame_height - 1) {
+        uint8_t *buf_ptr1 = ref_frame + y0 * pre_buf->stride + x0;
+        // Extend the border.
+        build_mc_border(buf_ptr1, pre_buf->stride, xd->mc_buf, x1 - x0 + 1,
+                        x0, y0, x1 - x0 + 1, y1 - y0 + 1, frame_width,
+                        frame_height);
+        buf_stride = x1 - x0 + 1;
+        buf_ptr = xd->mc_buf + y_pad * 3 * buf_stride + x_pad * 3;
+      }
+    } else {
+      // Wait until reference block is ready. Pad 7 more pixels as last 7
+      // pixels of each superblock row can be changed by next superblock row.
+       if (pbi->frame_parallel_decode)
+         vp9_frameworker_wait(pbi->frame_worker_owner, ref_frame_buf,
+                              (y1 + 7) << (plane == 0 ? 0 : 1));
+    }
+
+    inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
+                    subpel_y, sf, w, h, ref, kernel, xs, ys);
+  }
+}
+
+void vp9_dec_build_inter_predictors_sb(VP9Decoder *const pbi, MACROBLOCKD *xd,
+                                       int mi_row, int mi_col,
+                                       BLOCK_SIZE bsize) {
+  int plane;
+  const int mi_x = mi_col * MI_SIZE;
+  const int mi_y = mi_row * MI_SIZE;
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize,
+                                                        &xd->plane[plane]);
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+    const int bw = 4 * num_4x4_w;
+    const int bh = 4 * num_4x4_h;
+
+    if (xd->mi[0]->mbmi.sb_type < BLOCK_8X8) {
+      int i = 0, x, y;
+      assert(bsize == BLOCK_8X8);
+      for (y = 0; y < num_4x4_h; ++y)
+        for (x = 0; x < num_4x4_w; ++x)
+          dec_build_inter_predictors(pbi, xd, plane, i++, bw, bh,
+                                     4 * x, 4 * y, 4, 4, mi_x, mi_y);
+    } else {
+      dec_build_inter_predictors(pbi, xd, plane, 0, bw, bh,
+                                 0, 0, bw, bh, mi_x, mi_y);
+    }
+  }
+}

vp9/decoder/vp9_decodeframe.h

@@ -25,6 +25,9 @@ void vp9_decode_frame(struct VP9Decoder *pbi,
                       const uint8_t *data, const uint8_t *data_end,
                       const uint8_t **p_data_end);
 
+void vp9_dec_build_inter_predictors_sb(struct VP9Decoder *const pbi,
+                                       MACROBLOCKD *xd, int mi_row, int mi_col,
+                                       BLOCK_SIZE bsize);
 #ifdef __cplusplus
 }  // extern "C"
 #endif

vp9/decoder/vp9_decodemv.c

@@ -96,7 +96,7 @@ static void set_segment_id(VP9_COMMON *cm, BLOCK_SIZE bsize,
 
   for (y = 0; y < ymis; y++)
     for (x = 0; x < xmis; x++)
-      cm->last_frame_seg_map[mi_offset + y * cm->mi_cols + x] = segment_id;
+      cm->current_frame_seg_map[mi_offset + y * cm->mi_cols + x] = segment_id;
 }
 
 static int read_intra_segment_id(VP9_COMMON *const cm, MACROBLOCKD *const xd,
@@ -129,8 +129,10 @@ static int read_inter_segment_id(VP9_COMMON *const cm, MACROBLOCKD *const xd,
 
   predicted_segment_id = vp9_get_segment_id(cm, cm->last_frame_seg_map,
                                             bsize, mi_row, mi_col);
-  if (!seg->update_map)
+  if (!seg->update_map) {
+    set_segment_id(cm, bsize, mi_row, mi_col, predicted_segment_id);
     return predicted_segment_id;
+  }
 
   if (seg->temporal_update) {
     const vp9_prob pred_prob = vp9_get_pred_prob_seg_id(seg, xd);
@@ -418,11 +420,18 @@ static int read_is_inter_block(VP9_COMMON *const cm, MACROBLOCKD *const xd,
   }
 }
 
-static void read_inter_block_mode_info(VP9_COMMON *const cm,
+static void fpm_sync(void *const data, int mi_row) {
+  VP9Decoder *const pbi = (VP9Decoder *)data;
+  vp9_frameworker_wait(pbi->frame_worker_owner, pbi->prev_buf,
+                       mi_row << MI_BLOCK_SIZE_LOG2);
+}
+
+static void read_inter_block_mode_info(VP9Decoder *const pbi,
                                        MACROBLOCKD *const xd,
                                        const TileInfo *const tile,
                                        MODE_INFO *const mi,
                                        int mi_row, int mi_col, vp9_reader *r) {
+  VP9_COMMON *const cm = &pbi->common;
   MB_MODE_INFO *const mbmi = &mi->mbmi;
   const BLOCK_SIZE bsize = mbmi->sb_type;
   const int allow_hp = cm->allow_high_precision_mv;
@@ -436,7 +445,7 @@ static void read_inter_block_mode_info(VP9_COMMON *const cm,
   for (ref = 0; ref < 1 + is_compound; ++ref) {
     const MV_REFERENCE_FRAME frame = mbmi->ref_frame[ref];
     vp9_find_mv_refs(cm, xd, tile, mi, frame, mbmi->ref_mvs[frame],
-                     mi_row, mi_col);
+                     mi_row, mi_col, fpm_sync, (void *)pbi);
   }
 
   inter_mode_ctx = mbmi->mode_context[mbmi->ref_frame[0]];
@@ -510,10 +519,13 @@ static void read_inter_block_mode_info(VP9_COMMON *const cm,
   }
 }
 
-static void read_inter_frame_mode_info(VP9_COMMON *const cm,
+// TODO(hkuang): Pass cm instead of pbi. This requires change in
+// vp9_frameworker_wait.
+static void read_inter_frame_mode_info(VP9Decoder *const pbi,
                                        MACROBLOCKD *const xd,
                                        const TileInfo *const tile,
                                        int mi_row, int mi_col, vp9_reader *r) {
+  VP9_COMMON *const cm = &pbi->common;
   MODE_INFO *const mi = xd->mi[0];
   MB_MODE_INFO *const mbmi = &mi->mbmi;
   int inter_block;
@@ -527,16 +539,17 @@ static void read_inter_frame_mode_info(VP9_COMMON *const cm,
                                !mbmi->skip || !inter_block, r);
 
   if (inter_block)
-    read_inter_block_mode_info(cm, xd, tile, mi, mi_row, mi_col, r);
+    read_inter_block_mode_info(pbi, xd, tile, mi, mi_row, mi_col, r);
   else
     read_intra_block_mode_info(cm, mi, r);
 }
 
-void vp9_read_mode_info(VP9_COMMON *cm, MACROBLOCKD *xd,
+void vp9_read_mode_info(VP9Decoder *const pbi, MACROBLOCKD *xd,
                         const TileInfo *const tile,
                         int mi_row, int mi_col, vp9_reader *r) {
+  VP9_COMMON *const cm = &pbi->common;
   if (frame_is_intra_only(cm))
     read_intra_frame_mode_info(cm, xd, mi_row, mi_col, r);
   else
-    read_inter_frame_mode_info(cm, xd, tile, mi_row, mi_col, r);
+    read_inter_frame_mode_info(pbi, xd, tile, mi_row, mi_col, r);
 }

vp9/decoder/vp9_decodemv.h

@@ -11,6 +11,7 @@
 #ifndef VP9_DECODER_VP9_DECODEMV_H_
 #define VP9_DECODER_VP9_DECODEMV_H_
 
+#include "vp9/decoder/vp9_decoder.h"
 #include "vp9/decoder/vp9_reader.h"
 
 #ifdef __cplusplus
@@ -19,7 +20,7 @@ extern "C" {
 
 struct TileInfo;
 
-void vp9_read_mode_info(VP9_COMMON *cm, MACROBLOCKD *xd,
+void vp9_read_mode_info(VP9Decoder *const pbi, MACROBLOCKD *xd,
                         const struct TileInfo *const tile,
                         int mi_row, int mi_col, vp9_reader *r);
 

vp9/decoder/vp9_decoder.c

@@ -26,6 +26,7 @@
 #endif
 #include "vp9/common/vp9_quant_common.h"
 #include "vp9/common/vp9_systemdependent.h"
+#include "vp9/common/vp9_thread.h"
 
 #include "vp9/decoder/vp9_decodeframe.h"
 #include "vp9/decoder/vp9_decoder.h"
@@ -41,7 +42,7 @@ static void initialize_dec() {
   }
 }
 
-VP9Decoder *vp9_decoder_create() {
+VP9Decoder *vp9_decoder_create(BufferPool *const pool) {
   VP9Decoder *const pbi = vpx_memalign(32, sizeof(*pbi));
   VP9_COMMON *const cm = pbi ? &pbi->common : NULL;
 
@@ -57,15 +58,18 @@ VP9Decoder *vp9_decoder_create() {
   }
 
   cm->error.setjmp = 1;
+  pbi->need_resync = 1;
   initialize_dec();
 
   vp9_rtcd();
 
   // Initialize the references to not point to any frame buffers.
   vpx_memset(&cm->ref_frame_map, -1, sizeof(cm->ref_frame_map));
+  vpx_memset(&cm->next_ref_frame_map, -1, sizeof(cm->next_ref_frame_map));
 
   cm->current_video_frame = 0;
   pbi->ready_for_new_data = 1;
+  pbi->common.buffer_pool = pool;
 
   // vp9_init_dequantizer() is first called here. Add check in
   // frame_init_dequantizer() to avoid unnecessary calling of
@@ -124,7 +128,7 @@ vpx_codec_err_t vp9_copy_reference_dec(VP9Decoder *pbi,
    */
   if (ref_frame_flag == VP9_LAST_FLAG) {
     const YV12_BUFFER_CONFIG *const cfg =
-        &cm->frame_bufs[cm->ref_frame_map[0]].buf;
+        &cm->buffer_pool->frame_bufs[cm->ref_frame_map[0]].buf;
     if (!equal_dimensions(cfg, sd))
       vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
                          "Incorrect buffer dimensions");
@@ -143,6 +147,7 @@ vpx_codec_err_t vp9_set_reference_dec(VP9_COMMON *cm,
                                       VP9_REFFRAME ref_frame_flag,
                                       YV12_BUFFER_CONFIG *sd) {
   RefBuffer *ref_buf = NULL;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
 
   // TODO(jkoleszar): The decoder doesn't have any real knowledge of what the
   // encoder is using the frame buffers for. This is just a stub to keep the
@@ -170,11 +175,11 @@ vpx_codec_err_t vp9_set_reference_dec(VP9_COMMON *cm,
     const int free_fb = get_free_fb(cm);
     // Decrease ref_count since it will be increased again in
     // ref_cnt_fb() below.
-    cm->frame_bufs[free_fb].ref_count--;
+    --frame_bufs[free_fb].ref_count;
 
     // Manage the reference counters and copy image.
-    ref_cnt_fb(cm->frame_bufs, ref_fb_ptr, free_fb);
-    ref_buf->buf = &cm->frame_bufs[*ref_fb_ptr].buf;
+    ref_cnt_fb(frame_bufs, ref_fb_ptr, free_fb);
+    ref_buf->buf = &frame_bufs[*ref_fb_ptr].buf;
     vp8_yv12_copy_frame(sd, ref_buf->buf);
   }
 
@@ -184,11 +189,12 @@ vpx_codec_err_t vp9_set_reference_dec(VP9_COMMON *cm,
 
 int vp9_get_reference_dec(VP9Decoder *pbi, int index, YV12_BUFFER_CONFIG **fb) {
   VP9_COMMON *cm = &pbi->common;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
 
   if (index < 0 || index >= REF_FRAMES)
     return -1;
 
-  *fb = &cm->frame_bufs[cm->ref_frame_map[index]].buf;
+  *fb = &frame_bufs[cm->ref_frame_map[index]].buf;
   return 0;
 }
 
@@ -196,21 +202,38 @@ int vp9_get_reference_dec(VP9Decoder *pbi, int index, YV12_BUFFER_CONFIG **fb) {
 static void swap_frame_buffers(VP9Decoder *pbi) {
   int ref_index = 0, mask;
   VP9_COMMON *const cm = &pbi->common;
+  BufferPool *const pool = cm->buffer_pool;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
 
+  lock_buffer_pool(pool);
   for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
-    if (mask & 1) {
     const int old_idx = cm->ref_frame_map[ref_index];
-      ref_cnt_fb(cm->frame_bufs, &cm->ref_frame_map[ref_index],
-                 cm->new_fb_idx);
-      if (old_idx >= 0 && cm->frame_bufs[old_idx].ref_count == 0)
-        cm->release_fb_cb(cm->cb_priv,
-                          &cm->frame_bufs[old_idx].raw_frame_buffer);
+    // Current thread releases the holding of reference frame.
+    decrease_ref_count(old_idx, frame_bufs, pool);
+
+    // Release the reference frame in reference map.
+    if ((mask & 1) && old_idx >= 0) {
+      decrease_ref_count(old_idx, frame_bufs, pool);
     }
+    cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
     ++ref_index;
   }
 
+  // Current thread releases the holding of reference frame.
+  for (; ref_index < REF_FRAMES && !cm->show_existing_frame; ++ref_index) {
+    const int old_idx = cm->ref_frame_map[ref_index];
+    decrease_ref_count(old_idx, frame_bufs, pool);
+    cm->ref_frame_map[ref_index] = cm->next_ref_frame_map[ref_index];
+  }
+  unlock_buffer_pool(pool);
+  pbi->hold_ref_buf = 0;
   cm->frame_to_show = get_frame_new_buffer(cm);
-  cm->frame_bufs[cm->new_fb_idx].ref_count--;
+
+  if (!pbi->frame_parallel_decode || !cm->show_frame) {
+    lock_buffer_pool(pool);
+    --frame_bufs[cm->new_fb_idx].ref_count;
+    unlock_buffer_pool(pool);
+  }
 
   // Invalidate these references until the next frame starts.
   for (ref_index = 0; ref_index < 3; ref_index++)
@@ -220,9 +243,10 @@ static void swap_frame_buffers(VP9Decoder *pbi) {
 int vp9_receive_compressed_data(VP9Decoder *pbi,
                                 size_t size, const uint8_t **psource) {
   VP9_COMMON *const cm = &pbi->common;
+  BufferPool *const pool = cm->buffer_pool;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
   const uint8_t *source = *psource;
   int retcode = 0;
-
   cm->error.error_code = VPX_CODEC_OK;
 
   if (size == 0) {
@@ -238,27 +262,63 @@ int vp9_receive_compressed_data(VP9Decoder *pbi,
       cm->frame_refs[0].buf->corrupted = 1;
   }
 
+  pbi->ready_for_new_data = 0;
+
   // Check if the previous frame was a frame without any references to it.
-  if (cm->new_fb_idx >= 0 && cm->frame_bufs[cm->new_fb_idx].ref_count == 0)
-    cm->release_fb_cb(cm->cb_priv,
-                      &cm->frame_bufs[cm->new_fb_idx].raw_frame_buffer);
+  // Release frame buffer if not decoding in frame parallel mode.
+  if (!pbi->frame_parallel_decode && cm->new_fb_idx >= 0
+      && frame_bufs[cm->new_fb_idx].ref_count == 0)
+    pool->release_fb_cb(pool->cb_priv,
+                        &frame_bufs[cm->new_fb_idx].raw_frame_buffer);
   cm->new_fb_idx = get_free_fb(cm);
 
+  pbi->hold_ref_buf = 0;
+  if (pbi->frame_parallel_decode) {
+    VP9Worker *const worker = pbi->frame_worker_owner;
+    vp9_frameworker_lock_stats(worker);
+    frame_bufs[cm->new_fb_idx].frame_worker_owner = worker;
+    // Reset decoding progress.
+    pbi->cur_buf = &frame_bufs[cm->new_fb_idx];
+    pbi->cur_buf->row = -1;
+    pbi->cur_buf->col = -1;
+    vp9_frameworker_unlock_stats(worker);
+  } else {
+    pbi->cur_buf = &frame_bufs[cm->new_fb_idx];
+  }
+
   if (setjmp(cm->error.jmp)) {
     cm->error.setjmp = 0;
+    pbi->ready_for_new_data = 1;
 
-    // We do not know if the missing frame(s) was supposed to update
-    // any of the reference buffers, but we act conservative and
-    // mark only the last buffer as corrupted.
-    //
-    // TODO(jkoleszar): Error concealment is undefined and non-normative
-    // at this point, but if it becomes so, [0] may not always be the correct
-    // thing to do here.
-    if (cm->frame_refs[0].idx != INT_MAX && cm->frame_refs[0].buf != NULL)
-      cm->frame_refs[0].buf->corrupted = 1;
+    lock_buffer_pool(pool);
+    // Release all the reference buffers if worker thread is holding them.
+    if (pbi->hold_ref_buf == 1) {
+      int ref_index = 0, mask;
+      VP9_COMMON *const cm = &pbi->common;
+      BufferPool *const pool = cm->buffer_pool;
+      RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+      for (mask = pbi->refresh_frame_flags; mask; mask >>= 1) {
+        const int old_idx = cm->ref_frame_map[ref_index];
+        // Current thread releases the holding of reference frame.
+        decrease_ref_count(old_idx, frame_bufs, pool);
 
-    if (cm->new_fb_idx > 0 && cm->frame_bufs[cm->new_fb_idx].ref_count > 0)
-      cm->frame_bufs[cm->new_fb_idx].ref_count--;
+        // Release the reference frame in reference map.
+        if ((mask & 1) && old_idx >= 0) {
+          decrease_ref_count(old_idx, frame_bufs, pool);
+        }
+        ++ref_index;
+      }
+
+      // Current thread releases the holding of reference frame.
+      for (; ref_index < REF_FRAMES && !cm->show_existing_frame; ++ref_index) {
+        const int old_idx = cm->ref_frame_map[ref_index];
+        decrease_ref_count(old_idx, frame_bufs, pool);
+      }
+      pbi->hold_ref_buf = 0;
+    }
+    // Release current frame.
+    decrease_ref_count(cm->new_fb_idx, frame_bufs, pool);
+    unlock_buffer_pool(pool);
 
     return -1;
   }
@@ -271,19 +331,38 @@ int vp9_receive_compressed_data(VP9Decoder *pbi,
 
   vp9_clear_system_state();
 
-  cm->last_width = cm->width;
-  cm->last_height = cm->height;
-
   if (!cm->show_existing_frame)
     cm->last_show_frame = cm->show_frame;
+
+  // Update progress in frame parallel decode.
+  if (pbi->frame_parallel_decode) {
+    // Need to lock the mutex here as another thread may
+    // be accessing this buffer.
+    VP9Worker *const worker = pbi->frame_worker_owner;
+    FrameWorkerData *const frame_worker_data = worker->data1;
+    vp9_frameworker_lock_stats(worker);
+
+    if (cm->show_frame) {
+      if (!cm->show_existing_frame)
+        vp9_swap_mi_and_prev_mi(cm);
+      cm->current_video_frame++;
+    }
+    vp9_swap_current_and_last_seg_map(cm);
+    frame_worker_data->frame_decoded = 1;
+    frame_worker_data->frame_context_ready = 1;
+    vp9_frameworker_signal_stats(worker);
+    vp9_frameworker_unlock_stats(worker);
+  } else {
+    cm->last_width = cm->width;
+    cm->last_height = cm->height;
     if (cm->show_frame) {
       if (!cm->show_existing_frame)
         vp9_swap_mi_and_prev_mi(cm);
-
       cm->current_video_frame++;
     }
 
-  pbi->ready_for_new_data = 0;
+    vp9_swap_current_and_last_seg_map(cm);
+  }
 
   cm->error.setjmp = 0;
   return retcode;
@@ -299,12 +378,12 @@ int vp9_get_raw_frame(VP9Decoder *pbi, YV12_BUFFER_CONFIG *sd,
   if (pbi->ready_for_new_data == 1)
     return ret;
 
+  pbi->ready_for_new_data = 1;
+
   /* no raw frame to show!!! */
   if (pbi->common.show_frame == 0)
     return ret;
 
-  pbi->ready_for_new_data = 1;
-
 #if CONFIG_VP9_POSTPROC
   ret = vp9_post_proc_frame(&pbi->common, sd, flags);
 #else

vp9/decoder/vp9_decoder.h

@@ -18,10 +18,10 @@
 
 #include "vp9/common/vp9_onyxc_int.h"
 #include "vp9/common/vp9_ppflags.h"
+#include "vp9/common/vp9_thread.h"
 
 #include "vp9/decoder/vp9_decoder.h"
 #include "vp9/decoder/vp9_dthread.h"
-#include "vp9/decoder/vp9_thread.h"
 
 #ifdef __cplusplus
 extern "C" {
@@ -45,6 +45,12 @@ typedef struct VP9Decoder {
 
   int frame_parallel_decode;  // frame-based threading.
 
+  // TODO(hkuang): Combine this with cur_buf in macroblockd as they are
+  // the same.
+  RefCntBuffer *cur_buf;   //  Current decoding frame buffer.
+  RefCntBuffer *prev_buf;  //  Previous decoding frame buffer.
+
+  VP9Worker *frame_worker_owner;   // frame_worker that owns this pbi.
   VP9Worker lf_worker;
   VP9Worker *tile_workers;
   int num_tile_workers;
@@ -59,6 +65,8 @@ typedef struct VP9Decoder {
 
   int max_threads;
   int inv_tile_order;
+  int need_resync;  // wait for key/intra-only frame.
+  int hold_ref_buf;  // hold the reference buffer.
 } VP9Decoder;
 
 int vp9_receive_compressed_data(struct VP9Decoder *pbi,
@@ -78,10 +86,25 @@ vpx_codec_err_t vp9_set_reference_dec(VP9_COMMON *cm,
 int vp9_get_reference_dec(struct VP9Decoder *pbi,
                           int index, YV12_BUFFER_CONFIG **fb);
 
-struct VP9Decoder *vp9_decoder_create();
+struct VP9Decoder *vp9_decoder_create(BufferPool *const pool);
 
 void vp9_decoder_remove(struct VP9Decoder *pbi);
 
+static INLINE void decrease_ref_count(int idx, RefCntBuffer *const frame_bufs,
+                                      BufferPool *const pool) {
+  if (idx >= 0) {
+    --frame_bufs[idx].ref_count;
+    // A worker may only get a free framebuffer index when calling get_free_fb.
+    // But the private buffer is not set up until finish decoding header.
+    // So any error happens during decoding header, the frame_bufs will not
+    // have valid priv buffer.
+    if (frame_bufs[idx].ref_count == 0 &&
+        frame_bufs[idx].raw_frame_buffer.priv) {
+      pool->release_fb_cb(pool->cb_priv, &frame_bufs[idx].raw_frame_buffer);
+    }
+  }
+}
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif

vp9/decoder/vp9_dthread.c

@@ -17,6 +17,8 @@
 #include "vp9/decoder/vp9_dthread.h"
 #include "vp9/decoder/vp9_decoder.h"
 
+// #define DEBUG_THREAD
+
 #if CONFIG_MULTITHREAD
 static INLINE void mutex_lock(pthread_mutex_t *const mutex) {
   const int kMaxTryLocks = 4000;
@@ -279,3 +281,178 @@ void vp9_loop_filter_dealloc(VP9LfSync *lf_sync, int rows) {
     vp9_zero(*lf_sync);
   }
 }
+
+// TODO(hkuang): Clean up all the #ifdef in this file.
+void vp9_frameworker_lock_stats(VP9Worker *const worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const worker_data = worker->data1;
+  pthread_mutex_lock(&worker_data->stats_mutex);
+#else
+  (void)worker;
+#endif
+}
+
+void vp9_frameworker_unlock_stats(VP9Worker *const worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const worker_data = worker->data1;
+  pthread_mutex_unlock(&worker_data->stats_mutex);
+#else
+  (void)worker;
+#endif
+}
+
+void vp9_frameworker_signal_stats(VP9Worker *const worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const worker_data = worker->data1;
+  // TODO(hkuang): Investigate using broadcast or signal.
+  pthread_cond_signal(&worker_data->stats_cond);
+#else
+  (void)worker;
+#endif
+}
+
+// TODO(hkuang): Remove worker parameter as it is only used in debug code.
+void vp9_frameworker_wait(VP9Worker *const worker, RefCntBuffer *const ref_buf,
+                          int row) {
+#if CONFIG_MULTITHREAD
+  if (!ref_buf)
+    return;
+
+  // Enabling the following line of code will get harmless tsan error but
+  // will get best performance.
+  // if (ref_buf->row >= row && ref_buf->buf.corrupted != 1) return;
+
+  {
+    // Find the worker thread that owns the reference frame. If the reference
+    // frame has been fully decoded, it may not have owner.
+    VP9Worker *const ref_worker = ref_buf->frame_worker_owner;
+    FrameWorkerData *const ref_worker_data =
+        (FrameWorkerData *)ref_worker->data1;
+    const VP9Decoder *const pbi = ref_worker_data->pbi;
+
+#ifdef DEBUG_THREAD
+    {
+      FrameWorkerData *const worker_data = (FrameWorkerData *)worker->data1;
+      printf("%d %p worker is waiting for %d %p worker (%d)  ref %d \r\n",
+             worker_data->worker_id, worker, ref_worker_data->worker_id,
+             ref_buf->frame_worker_owner, row, ref_buf->row);
+    }
+#endif
+
+    vp9_frameworker_lock_stats(ref_worker);
+    while (ref_buf->row < row && pbi->cur_buf == ref_buf &&
+           ref_buf->buf.corrupted != 1) {
+      pthread_cond_wait(&ref_worker_data->stats_cond,
+                        &ref_worker_data->stats_mutex);
+    }
+
+    if (ref_buf->buf.corrupted == 1) {
+      FrameWorkerData *const worker_data = (FrameWorkerData *)worker->data1;
+      vp9_frameworker_unlock_stats(ref_worker);
+      vpx_internal_error(&worker_data->pbi->common.error,
+                         VPX_CODEC_CORRUPT_FRAME,
+                         "Worker %p failed to decode frame", worker);
+    }
+    vp9_frameworker_unlock_stats(ref_worker);
+  }
+#else
+  (void)ref_buf;
+  (void)row;
+  (void)ref_buf;
+#endif  // CONFIG_MULTITHREAD
+}
+
+void vp9_frameworker_broadcast(RefCntBuffer *const buf, int row) {
+#if CONFIG_MULTITHREAD
+  VP9Worker *worker = buf->frame_worker_owner;
+
+#ifdef DEBUG_THREAD
+  {
+    FrameWorkerData *const worker_data = (FrameWorkerData *)worker->data1;
+    printf("%d %p worker decode to (%d) \r\n", worker_data->worker_id,
+           buf->frame_worker_owner, row);
+  }
+#endif
+
+  vp9_frameworker_lock_stats(worker);
+  buf->row = row;
+  vp9_frameworker_signal_stats(worker);
+  vp9_frameworker_unlock_stats(worker);
+#else
+  (void)buf;
+  (void)row;
+#endif  // CONFIG_MULTITHREAD
+}
+
+void vp9_frameworker_copy_context(VP9Worker *const dst_worker,
+                                  VP9Worker *const src_worker) {
+#if CONFIG_MULTITHREAD
+  FrameWorkerData *const src_worker_data = (FrameWorkerData *)src_worker->data1;
+  FrameWorkerData *const dst_worker_data = (FrameWorkerData *)dst_worker->data1;
+  VP9_COMMON *const src_cm = &src_worker_data->pbi->common;
+  VP9_COMMON *const dst_cm = &dst_worker_data->pbi->common;
+  int i;
+
+  // Wait until source frame's context is ready.
+  vp9_frameworker_lock_stats(src_worker);
+  while (!src_worker_data->frame_context_ready) {
+    pthread_cond_wait(&src_worker_data->stats_cond,
+        &src_worker_data->stats_mutex);
+  }
+
+  // src worker may have already finished decoding a frame and swapped the mi.
+  // TODO(hkuang): Remove following code after implenment no ModeInfo decoding.
+  if (src_worker_data->frame_decoded) {
+    dst_cm->prev_mip = src_cm->prev_mip;
+    dst_cm->prev_mi = src_cm->prev_mi;
+    dst_cm->prev_mi_grid_base = src_cm->prev_mi_grid_base;
+    dst_cm->prev_mi_grid_visible = src_cm->prev_mi_grid_visible;
+    dst_cm->last_frame_seg_map = src_cm->last_frame_seg_map;
+  } else {
+    dst_cm->prev_mip = src_cm->mip;
+    dst_cm->prev_mi = src_cm->mi;
+    dst_cm->prev_mi_grid_base = src_cm->mi_grid_base;
+    dst_cm->prev_mi_grid_visible = src_cm->mi_grid_visible;
+    dst_cm->last_frame_seg_map = src_cm->current_frame_seg_map;
+  }
+  dst_worker_data->pbi->need_resync = src_worker_data->pbi->need_resync;
+  vp9_frameworker_unlock_stats(src_worker);
+
+  dst_worker_data->pbi->prev_buf =
+      src_worker_data->pbi->common.show_existing_frame ?
+          NULL : src_worker_data->pbi->cur_buf;
+
+  dst_cm->last_width = !src_cm->show_existing_frame ?
+                       src_cm->width : src_cm->last_width;
+  dst_cm->last_height = !src_cm->show_existing_frame ?
+                        src_cm->height : src_cm->last_height;
+  dst_cm->display_width = src_cm->display_width;
+  dst_cm->display_height = src_cm->display_height;
+  dst_cm->subsampling_x = src_cm->subsampling_x;
+  dst_cm->subsampling_y = src_cm->subsampling_y;
+  dst_cm->last_show_frame = !src_cm->show_existing_frame ?
+                            src_cm->show_frame : src_cm->last_show_frame;
+  dst_cm->last_frame_type = src_cm->last_frame_type;
+  dst_cm->frame_type = src_cm->frame_type;
+  dst_cm->y_dc_delta_q = src_cm->y_dc_delta_q;
+  dst_cm->uv_dc_delta_q = src_cm->uv_dc_delta_q;
+  dst_cm->uv_ac_delta_q = src_cm->uv_ac_delta_q;
+  dst_cm->base_qindex = src_cm->base_qindex;
+
+  for (i = 0; i < REF_FRAMES; ++i)
+    dst_cm->ref_frame_map[i] = src_cm->next_ref_frame_map[i];
+
+  memcpy(dst_cm->lf_info.lfthr, src_cm->lf_info.lfthr,
+         (MAX_LOOP_FILTER + 1) * sizeof(loop_filter_thresh));
+  dst_cm->lf.last_sharpness_level = src_cm->lf.sharpness_level;
+  dst_cm->lf.filter_level = src_cm->lf.filter_level;
+  memcpy(dst_cm->lf.ref_deltas, src_cm->lf.ref_deltas, MAX_REF_LF_DELTAS);
+  memcpy(dst_cm->lf.mode_deltas, src_cm->lf.mode_deltas, MAX_MODE_LF_DELTAS);
+  dst_cm->seg = src_cm->seg;
+  memcpy(dst_cm->frame_contexts, src_cm->frame_contexts,
+         FRAME_CONTEXTS * sizeof(dst_cm->frame_contexts[0]));
+#else
+  (void) dst_worker;
+  (void) src_worker;
+#endif  // CONFIG_MULTITHREAD
+}

vp9/decoder/vp9_dthread.h

@@ -12,14 +12,14 @@
 #define VP9_DECODER_VP9_DTHREAD_H_
 
 #include "./vpx_config.h"
+#include "vp9/common/vp9_thread.h"
 #include "vp9/decoder/vp9_reader.h"
-#include "vp9/decoder/vp9_thread.h"
 
 struct VP9Common;
 struct VP9Decoder;
 
 typedef struct TileWorkerData {
-  struct VP9Common *cm;
+  struct VP9Decoder *pbi;
   vp9_reader bit_reader;
   DECLARE_ALIGNED(16, struct macroblockd, xd);
 
@@ -40,6 +40,31 @@ typedef struct VP9LfSyncData {
   int sync_range;
 } VP9LfSync;
 
+// WorkerData for the FrameWorker thread. It contains all the information of
+// the worker and decode structures for decoding a frame.
+typedef struct FrameWorkerData {
+  struct VP9Decoder *pbi;
+  const uint8_t *data;
+  const uint8_t *data_end;
+  size_t data_size;
+  void *user_priv;
+  int result;
+  int worker_id;
+
+  // scratch_buffer is used in frame parallel mode only.
+  // It is used to make a copy of the compressed data.
+  uint8_t *scratch_buffer;
+  size_t scratch_buffer_size;
+
+#if CONFIG_MULTITHREAD
+  pthread_mutex_t stats_mutex;
+  pthread_cond_t stats_cond;
+#endif
+
+  int frame_context_ready;  // Current frame's context is ready to read.
+  int frame_decoded;        // Finished decoding current frame.
+} FrameWorkerData;
+
 // Allocate memory for loopfilter row synchronization.
 void vp9_loop_filter_alloc(struct VP9Common *cm, VP9LfSync *lf_sync,
                            int rows, int width);
@@ -54,4 +79,23 @@ void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame,
                               int frame_filter_level,
                               int y_only);
 
+void vp9_frameworker_lock_stats(VP9Worker *const worker);
+void vp9_frameworker_unlock_stats(VP9Worker *const worker);
+void vp9_frameworker_signal_stats(VP9Worker *const worker);
+
+// Wait until ref_buf has been decoded to row in real pixel unit.
+// Note: worker may already finish decoding ref_buf and release it in order to
+// start decoding next frame. So need to check whether worker is still decoding
+// ref_buf.
+void vp9_frameworker_wait(VP9Worker *const worker, RefCntBuffer *const ref_buf,
+                          int row);
+
+// FrameWorker broadcasts its decoding progress so other workers that are
+// waiting on it can resume decoding.
+void vp9_frameworker_broadcast(RefCntBuffer *const buf, int row);
+
+// Copy necessary decoding context from src worker to dst worker.
+void vp9_frameworker_copy_context(VP9Worker *const dst_worker,
+                                  VP9Worker *const src_worker);
+
 #endif  // VP9_DECODER_VP9_DTHREAD_H_

vp9/encoder/vp9_encoder.c

@@ -713,7 +713,8 @@ static void cal_nmvsadcosts_hp(int *mvsadcost[2]) {
 }
 
 
-VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf) {
+VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
+                                BufferPool *const pool) {
   unsigned int i, j;
   VP9_COMP *const cpi = vpx_memalign(32, sizeof(VP9_COMP));
   VP9_COMMON *const cm = cpi != NULL ? &cpi->common : NULL;
@@ -734,6 +735,7 @@ VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf) {
   vp9_rtcd();
 
   cpi->use_svc = 0;
+  cpi->common.buffer_pool = pool;
 
   init_config(cpi, oxcf);
   vp9_rc_init(&cpi->oxcf, cpi->pass, &cpi->rc);
@@ -1273,7 +1275,7 @@ int vp9_get_reference_enc(VP9_COMP *cpi, int index, YV12_BUFFER_CONFIG **fb) {
   if (index < 0 || index >= REF_FRAMES)
     return -1;
 
-  *fb = &cm->frame_bufs[cm->ref_frame_map[index]].buf;
+  *fb = &cm->buffer_pool->frame_bufs[cm->ref_frame_map[index]].buf;
   return 0;
 }
 
@@ -1542,14 +1544,13 @@ static int recode_loop_test(const VP9_COMP *cpi,
 
 void vp9_update_reference_frames(VP9_COMP *cpi) {
   VP9_COMMON * const cm = &cpi->common;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
 
   // At this point the new frame has been encoded.
   // If any buffer copy / swapping is signaled it should be done here.
   if (cm->frame_type == KEY_FRAME) {
-    ref_cnt_fb(cm->frame_bufs,
-               &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
-    ref_cnt_fb(cm->frame_bufs,
-               &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
+    ref_cnt_fb(frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
+    ref_cnt_fb(frame_bufs, &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
   } else if (!cpi->multi_arf_allowed && cpi->refresh_golden_frame &&
              cpi->rc.is_src_frame_alt_ref && !cpi->use_svc) {
     /* Preserve the previously existing golden frame and update the frame in
@@ -1563,8 +1564,7 @@ void vp9_update_reference_frames(VP9_COMP *cpi) {
      */
     int tmp;
 
-    ref_cnt_fb(cm->frame_bufs,
-               &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
+    ref_cnt_fb(frame_bufs, &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx);
 
     tmp = cpi->alt_fb_idx;
     cpi->alt_fb_idx = cpi->gld_fb_idx;
@@ -1577,19 +1577,17 @@ void vp9_update_reference_frames(VP9_COMP *cpi) {
         arf_idx = gf_group->arf_update_idx[gf_group->index];
       }
 
-      ref_cnt_fb(cm->frame_bufs,
-                 &cm->ref_frame_map[arf_idx], cm->new_fb_idx);
+      ref_cnt_fb(frame_bufs, &cm->ref_frame_map[arf_idx], cm->new_fb_idx);
     }
 
     if (cpi->refresh_golden_frame) {
-      ref_cnt_fb(cm->frame_bufs,
+      ref_cnt_fb(frame_bufs,
                  &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx);
     }
   }
 
   if (cpi->refresh_last_frame) {
-    ref_cnt_fb(cm->frame_bufs,
-               &cm->ref_frame_map[cpi->lst_fb_idx], cm->new_fb_idx);
+    ref_cnt_fb(frame_bufs, &cm->ref_frame_map[cpi->lst_fb_idx], cm->new_fb_idx);
   }
 #if CONFIG_DENOISING
   vp9_denoiser_update_frame_info(&cpi->denoiser,
@@ -1630,34 +1628,36 @@ void vp9_scale_references(VP9_COMP *cpi) {
   VP9_COMMON *cm = &cpi->common;
   MV_REFERENCE_FRAME ref_frame;
   const VP9_REFFRAME ref_mask[3] = {VP9_LAST_FLAG, VP9_GOLD_FLAG, VP9_ALT_FLAG};
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
 
   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
     const int idx = cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)];
-    const YV12_BUFFER_CONFIG *const ref = &cm->frame_bufs[idx].buf;
+    const YV12_BUFFER_CONFIG *const ref = &frame_bufs[idx].buf;
 
     // Need to convert from VP9_REFFRAME to index into ref_mask (subtract 1).
     if ((cpi->ref_frame_flags & ref_mask[ref_frame - 1]) &&
         (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height)) {
       const int new_fb = get_free_fb(cm);
-      vp9_realloc_frame_buffer(&cm->frame_bufs[new_fb].buf,
+      vp9_realloc_frame_buffer(&frame_bufs[new_fb].buf,
                                cm->width, cm->height,
                                cm->subsampling_x, cm->subsampling_y,
                                VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL);
-      scale_and_extend_frame(ref, &cm->frame_bufs[new_fb].buf);
+      scale_and_extend_frame(ref, &frame_bufs[new_fb].buf);
       cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
     } else {
       cpi->scaled_ref_idx[ref_frame - 1] = idx;
-      cm->frame_bufs[idx].ref_count++;
+      ++frame_bufs[idx].ref_count;
     }
   }
 }
 
 static void release_scaled_references(VP9_COMP *cpi) {
   VP9_COMMON *cm = &cpi->common;
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
   int i;
 
-  for (i = 0; i < 3; i++)
-    cm->frame_bufs[cpi->scaled_ref_idx[i]].ref_count--;
+  for (i = 0; i < 3; ++i)
+    --frame_bufs[cpi->scaled_ref_idx[i]].ref_count;
 }
 
 static void full_to_model_count(unsigned int *model_count,
@@ -2520,6 +2520,7 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
   const int is_spatial_svc = cpi->use_svc &&
                              (cpi->svc.number_temporal_layers == 1) &&
                              (cpi->svc.number_spatial_layers > 1);
+  RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
 
   if (!cpi)
     return -1;
@@ -2602,7 +2603,8 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
       cm->show_frame = 1;
       cm->intra_only = 0;
 
-      // Check to see if the frame should be encoded as an arf overlay.
+      // Check to see if the frame to be encoded is an overlay for a previous
+      // arf frame and if so configure it as such.
       check_src_altref(cpi);
     }
   }
@@ -2656,7 +2658,7 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
   /* find a free buffer for the new frame, releasing the reference previously
    * held.
    */
-  cm->frame_bufs[cm->new_fb_idx].ref_count--;
+  --frame_bufs[cm->new_fb_idx].ref_count;
   cm->new_fb_idx = get_free_fb(cm);
 
   if (!cpi->use_svc && cpi->multi_arf_allowed) {
@@ -2690,7 +2692,7 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
 
   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
     const int idx = cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)];
-    YV12_BUFFER_CONFIG *const buf = &cm->frame_bufs[idx].buf;
+    YV12_BUFFER_CONFIG *const buf = &frame_bufs[idx].buf;
     RefBuffer *const ref_buf = &cm->frame_refs[ref_frame - 1];
     ref_buf->buf = buf;
     ref_buf->idx = idx;

vp9/encoder/vp9_encoder.h

@@ -436,7 +436,8 @@ typedef struct VP9_COMP {
 
 void vp9_initialize_enc();
 
-struct VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf);
+struct VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
+                                       BufferPool *const pool);
 void vp9_remove_compressor(VP9_COMP *cpi);
 
 void vp9_change_config(VP9_COMP *cpi, const VP9EncoderConfig *oxcf);
@@ -494,8 +495,9 @@ static INLINE int get_ref_frame_idx(const VP9_COMP *cpi,
 
 static INLINE YV12_BUFFER_CONFIG *get_ref_frame_buffer(
     VP9_COMP *cpi, MV_REFERENCE_FRAME ref_frame) {
-  VP9_COMMON * const cm = &cpi->common;
-  return &cm->frame_bufs[cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)]]
+  VP9_COMMON *const cm = &cpi->common;
+  BufferPool *const pool = cm->buffer_pool;
+  return &pool->frame_bufs[cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)]]
       .buf;
 }
 

vp9/encoder/vp9_pickmode.c

@@ -484,7 +484,7 @@ int64_t vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
 
       if (cm->coding_use_prev_mi)
         vp9_find_mv_refs(cm, xd, tile, xd->mi[0], ref_frame,
-                         candidates, mi_row, mi_col);
+                         candidates, mi_row, mi_col, NULL, NULL);
       else
         const_motion[ref_frame] = mv_refs_rt(cm, xd, tile, xd->mi[0],
                                              ref_frame, candidates,

vp9/encoder/vp9_rdopt.c

@@ -2253,7 +2253,8 @@ void vp9_setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
   vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
 
   // Gets an initial list of candidate vectors from neighbours and orders them
-  vp9_find_mv_refs(cm, xd, tile, mi, ref_frame, candidates, mi_row, mi_col);
+  vp9_find_mv_refs(cm, xd, tile, mi, ref_frame, candidates, mi_row, mi_col,
+                   NULL, NULL);
 
   // Candidate refinement carried out at encoder and decoder
   vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
@@ -2271,9 +2272,10 @@ void vp9_setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
 const YV12_BUFFER_CONFIG *vp9_get_scaled_ref_frame(const VP9_COMP *cpi,
                                                    int ref_frame) {
   const VP9_COMMON *const cm = &cpi->common;
+  const RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
   const int ref_idx = cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)];
   const int scaled_idx = cpi->scaled_ref_idx[ref_frame - 1];
-  return (scaled_idx != ref_idx) ? &cm->frame_bufs[scaled_idx].buf : NULL;
+  return (scaled_idx != ref_idx) ? &frame_bufs[scaled_idx].buf : NULL;
 }
 
 int vp9_get_switchable_rate(const VP9_COMP *cpi) {

vp9/vp9_common.mk

@@ -50,6 +50,8 @@ VP9_COMMON_SRCS-yes += common/vp9_seg_common.h
 VP9_COMMON_SRCS-yes += common/vp9_seg_common.c
 VP9_COMMON_SRCS-yes += common/vp9_systemdependent.h
 VP9_COMMON_SRCS-yes += common/vp9_textblit.h
+VP9_COMMON_SRCS-yes += common/vp9_thread.h
+VP9_COMMON_SRCS-yes += common/vp9_thread.c
 VP9_COMMON_SRCS-yes += common/vp9_tile_common.h
 VP9_COMMON_SRCS-yes += common/vp9_tile_common.c
 VP9_COMMON_SRCS-yes += common/vp9_loopfilter.c

vp9/vp9_cx_iface.c

@@ -90,6 +90,8 @@ struct vpx_codec_alg_priv {
   vp8_postproc_cfg_t      preview_ppcfg;
   vpx_codec_pkt_list_decl(128) pkt_list;
   unsigned int                 fixed_kf_cntr;
+  // BufferPool that holds all reference frames.
+  BufferPool              *buffer_pool;
 };
 
 static VP9_REFFRAME ref_frame_to_vp9_reframe(vpx_ref_frame_type_t frame) {
@@ -630,6 +632,10 @@ static vpx_codec_err_t encoder_init(vpx_codec_ctx_t *ctx,
     ctx->priv->alg_priv = priv;
     ctx->priv->init_flags = ctx->init_flags;
     ctx->priv->enc.total_encoders = 1;
+    ctx->priv->alg_priv->buffer_pool =
+        (BufferPool *)vpx_calloc(1, sizeof(BufferPool));
+    if (ctx->priv->alg_priv->buffer_pool == NULL)
+      return VPX_CODEC_MEM_ERROR;
 
     if (ctx->config.enc) {
       // Update the reference to the config structure to an internal copy.
@@ -667,7 +673,8 @@ static vpx_codec_err_t encoder_init(vpx_codec_ctx_t *ctx,
       set_encoder_config(&ctx->priv->alg_priv->oxcf,
                          &ctx->priv->alg_priv->cfg,
                          &ctx->priv->alg_priv->extra_cfg);
-      cpi = vp9_create_compressor(&ctx->priv->alg_priv->oxcf);
+      cpi = vp9_create_compressor(&ctx->priv->alg_priv->oxcf,
+                                  ctx->priv->alg_priv->buffer_pool);
       if (cpi == NULL)
         res = VPX_CODEC_MEM_ERROR;
       else
@@ -681,6 +688,7 @@ static vpx_codec_err_t encoder_init(vpx_codec_ctx_t *ctx,
 static vpx_codec_err_t encoder_destroy(vpx_codec_alg_priv_t *ctx) {
   free(ctx->cx_data);
   vp9_remove_compressor(ctx->cpi);
+  vpx_free(ctx->buffer_pool);
   free(ctx);
   return VPX_CODEC_OK;
 }

vp9/vp9_dx_iface.c

@@ -18,6 +18,7 @@
 #include "vpx/vpx_decoder.h"
 
 #include "vp9/common/vp9_frame_buffers.h"
+#include "vp9/common/vp9_thread.h"
 
 #include "vp9/decoder/vp9_decoder.h"
 #include "vp9/decoder/vp9_read_bit_buffer.h"
@@ -28,19 +29,43 @@
 
 typedef vpx_codec_stream_info_t vp9_stream_info_t;
 
+// This limit is due to framebuffer numbers.
+// TODO(hkuang): Remove this limit after implementing ondemand framebuffers.
+#define FRAME_CACHE_SIZE 6   // Cache maximum 6 decoded frames.
+
+typedef struct cache_frame {
+  int fb_idx;
+  vpx_image_t img;
+} cache_frame;
+
 struct vpx_codec_alg_priv {
   vpx_codec_priv_t        base;
   vpx_codec_dec_cfg_t     cfg;
   vp9_stream_info_t       si;
-  struct VP9Decoder *pbi;
   int                     postproc_cfg_set;
   vp8_postproc_cfg_t      postproc_cfg;
   vpx_decrypt_cb          decrypt_cb;
   void                    *decrypt_state;
   vpx_image_t             img;
-  int                     img_avail;
+  int                     flushed;
   int                     invert_tile_order;
+  int                     last_show_frame;  // Index of last output frame.
+
+  // Frame parallel related.
   int                     frame_parallel_decode;  // frame-based threading.
+  VP9Worker               *frame_workers;
+  int                     num_frame_workers;
+  int                     next_submit_worker_id;
+  int                     last_submit_worker_id;
+  int                     next_output_worker_id;
+  int                     available_threads;
+  cache_frame             frame_cache[FRAME_CACHE_SIZE];
+  int                     frame_cache_write;
+  int                     frame_cache_read;
+  int                     num_cache_frames;
+
+  // BufferPool that holds all reference frames. Shared by all the FrameWorkers.
+  BufferPool              *buffer_pool;
 
   // External frame buffer info to save for VP9 common.
   void *ext_priv;  // Private data associated with the external frame buffers.
@@ -68,11 +93,11 @@ static vpx_codec_err_t decoder_init(vpx_codec_ctx_t *ctx,
     ctx->priv->alg_priv = alg_priv;
     ctx->priv->alg_priv->si.sz = sizeof(ctx->priv->alg_priv->si);
     ctx->priv->init_flags = ctx->init_flags;
+    ctx->priv->alg_priv->flushed = 0;
+    // Only do frame parallel decode when threads > 1.
     ctx->priv->alg_priv->frame_parallel_decode =
-        (ctx->init_flags & VPX_CODEC_USE_FRAME_THREADING);
-
-    // Disable frame parallel decoding for now.
-    ctx->priv->alg_priv->frame_parallel_decode = 0;
+        ((ctx->config.dec->threads > 1) &&
+         (ctx->init_flags & VPX_CODEC_USE_FRAME_THREADING)) ? 1 : 0;
 
     if (ctx->config.dec) {
       // Update the reference to the config structure to an internal copy.
@@ -85,11 +110,29 @@ static vpx_codec_err_t decoder_init(vpx_codec_ctx_t *ctx,
 }
 
 static vpx_codec_err_t decoder_destroy(vpx_codec_alg_priv_t *ctx) {
-  if (ctx->pbi) {
-    vp9_decoder_remove(ctx->pbi);
-    ctx->pbi = NULL;
+  if (ctx->frame_workers != NULL) {
+    int i;
+    for (i = 0; i < ctx->num_frame_workers; ++i) {
+      VP9Worker *const worker = &ctx->frame_workers[i];
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      vp9_get_worker_interface()->end(worker);
+      vp9_decoder_remove(frame_worker_data->pbi);
+      vpx_free(frame_worker_data->scratch_buffer);
+#if CONFIG_MULTITHREAD
+      pthread_mutex_destroy(&frame_worker_data->stats_mutex);
+      pthread_cond_destroy(&frame_worker_data->stats_cond);
+#endif
+      vpx_free(frame_worker_data);
+    }
+#if CONFIG_MULTITHREAD
+    pthread_mutex_destroy(&ctx->buffer_pool->pool_mutex);
+#endif
+    vp9_free_internal_frame_buffers(&ctx->buffer_pool->int_frame_buffers);
   }
 
+  vpx_free(ctx->frame_workers);
+  vpx_free(ctx->buffer_pool);
   vpx_free(ctx);
 
   return VPX_CODEC_OK;
@@ -188,32 +231,43 @@ static vpx_codec_err_t decoder_get_si(vpx_codec_alg_priv_t *ctx,
   return VPX_CODEC_OK;
 }
 
+static void set_error_detail(vpx_codec_alg_priv_t *ctx,
+                             const char *const error) {
+  ctx->base.err_detail = error;
+}
+
 static vpx_codec_err_t update_error_state(vpx_codec_alg_priv_t *ctx,
                            const struct vpx_internal_error_info *error) {
   if (error->error_code)
-    ctx->base.err_detail = error->has_detail ? error->detail : NULL;
+    set_error_detail(ctx, error->has_detail ? error->detail : NULL);
 
   return error->error_code;
 }
 
 static void init_buffer_callbacks(vpx_codec_alg_priv_t *ctx) {
-  VP9_COMMON *const cm = &ctx->pbi->common;
+  int i;
+
+  for (i = 0; i < ctx->num_frame_workers; ++i) {
+    VP9Worker *const worker = &ctx->frame_workers[i];
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    VP9_COMMON *const cm = &frame_worker_data->pbi->common;
+    BufferPool *const pool = cm->buffer_pool;
 
     cm->new_fb_idx = -1;
-
     if (ctx->get_ext_fb_cb != NULL && ctx->release_ext_fb_cb != NULL) {
-    cm->get_fb_cb = ctx->get_ext_fb_cb;
-    cm->release_fb_cb = ctx->release_ext_fb_cb;
-    cm->cb_priv = ctx->ext_priv;
+      pool->get_fb_cb = ctx->get_ext_fb_cb;
+      pool->release_fb_cb = ctx->release_ext_fb_cb;
+      pool->cb_priv = ctx->ext_priv;
     } else {
-    cm->get_fb_cb = vp9_get_frame_buffer;
-    cm->release_fb_cb = vp9_release_frame_buffer;
+      pool->get_fb_cb = vp9_get_frame_buffer;
+      pool->release_fb_cb = vp9_release_frame_buffer;
 
-    if (vp9_alloc_internal_frame_buffers(&cm->int_frame_buffers))
+      if (vp9_alloc_internal_frame_buffers(&pool->int_frame_buffers))
         vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
                            "Failed to initialize internal frame buffers");
 
-    cm->cb_priv = &cm->int_frame_buffers;
+      pool->cb_priv = &pool->int_frame_buffers;
+    }
   }
 }
 
@@ -232,14 +286,126 @@ static void set_ppflags(const vpx_codec_alg_priv_t *ctx,
   flags->noise_level = ctx->postproc_cfg.noise_level;
 }
 
-static void init_decoder(vpx_codec_alg_priv_t *ctx) {
-  ctx->pbi = vp9_decoder_create();
-  if (ctx->pbi == NULL)
-    return;
+static int frame_worker_hook(void *arg1, void *arg2) {
+  FrameWorkerData *const frame_worker_data = (FrameWorkerData *)arg1;
+  const uint8_t *data = frame_worker_data->data;
+  (void)arg2;
 
-  ctx->pbi->max_threads = ctx->cfg.threads;
-  ctx->pbi->inv_tile_order = ctx->invert_tile_order;
-  ctx->pbi->frame_parallel_decode = ctx->frame_parallel_decode;
+  frame_worker_data->result =
+      vp9_receive_compressed_data(frame_worker_data->pbi,
+                                  frame_worker_data->data_size,
+                                  &data);
+  frame_worker_data->data_end = data;
+
+  if (frame_worker_data->pbi->frame_parallel_decode) {
+    // In frame parallel decoding, a worker thread must successfully decode all
+    // the compressed data.
+    if (frame_worker_data->result != 0 ||
+        frame_worker_data->data + frame_worker_data->data_size - 1 > data) {
+      VP9Worker *const worker = frame_worker_data->pbi->frame_worker_owner;
+      BufferPool *const pool = frame_worker_data->pbi->common.buffer_pool;
+      // Signal all the other threads that are waiting for this frame.
+      vp9_frameworker_lock_stats(worker);
+      frame_worker_data->frame_context_ready = 1;
+      lock_buffer_pool(pool);
+      frame_worker_data->pbi->cur_buf->buf.corrupted = 1;
+      unlock_buffer_pool(pool);
+      frame_worker_data->pbi->need_resync = 1;
+      vp9_frameworker_signal_stats(worker);
+      vp9_frameworker_unlock_stats(worker);
+      return 0;
+    }
+  } else if (frame_worker_data->result != 0) {
+    // Check decode result in serial decode.
+    frame_worker_data->pbi->cur_buf->buf.corrupted = 1;
+    frame_worker_data->pbi->need_resync = 1;
+  }
+
+  return !frame_worker_data->result;
+}
+
+static vpx_codec_err_t init_decoder(vpx_codec_alg_priv_t *ctx) {
+  int i;
+  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
+
+  ctx->last_show_frame = -1;
+  ctx->next_submit_worker_id = 0;
+  ctx->last_submit_worker_id = 0;
+  ctx->next_output_worker_id = 0;
+  ctx->frame_cache_read = 0;
+  ctx->frame_cache_write = 0;
+  ctx->num_cache_frames = 0;
+  ctx->num_frame_workers =
+      (ctx->frame_parallel_decode == 1) ? ctx->cfg.threads: 1;
+  ctx->available_threads = ctx->num_frame_workers;
+  ctx->flushed = 0;
+
+  ctx->buffer_pool = (BufferPool *)vpx_calloc(1, sizeof(BufferPool));
+  if (ctx->buffer_pool == NULL)
+    return VPX_CODEC_MEM_ERROR;
+
+#if CONFIG_MULTITHREAD
+    if (pthread_mutex_init(&ctx->buffer_pool->pool_mutex, NULL)) {
+      set_error_detail(ctx, "Failed to allocate buffer pool mutex");
+      return VPX_CODEC_MEM_ERROR;
+    }
+#endif
+
+  ctx->frame_workers = (VP9Worker *)
+      vpx_malloc(ctx->num_frame_workers * sizeof(*ctx->frame_workers));
+  if (ctx->frame_workers == NULL) {
+    set_error_detail(ctx, "Failed to allocate frame_workers");
+    return VPX_CODEC_MEM_ERROR;
+  }
+
+  for (i = 0; i < ctx->num_frame_workers; ++i) {
+    VP9Worker *const worker = &ctx->frame_workers[i];
+    FrameWorkerData *frame_worker_data = NULL;
+    winterface->init(worker);
+    worker->data1 = vpx_memalign(32, sizeof(FrameWorkerData));
+    if (worker->data1 == NULL) {
+      set_error_detail(ctx, "Failed to allocate frame_worker_data");
+      return VPX_CODEC_MEM_ERROR;
+    }
+    frame_worker_data = (FrameWorkerData *)worker->data1;
+    frame_worker_data->pbi = vp9_decoder_create(ctx->buffer_pool);
+    if (frame_worker_data->pbi == NULL) {
+      set_error_detail(ctx, "Failed to allocate frame_worker_data");
+      return VPX_CODEC_MEM_ERROR;
+    }
+    frame_worker_data->pbi->frame_worker_owner = worker;
+    frame_worker_data->pbi->common.mi_idx = 0;
+    frame_worker_data->pbi->common.prev_mi_idx = 1;
+    frame_worker_data->worker_id = i;
+    frame_worker_data->scratch_buffer = NULL;
+    frame_worker_data->scratch_buffer_size = 0;
+    frame_worker_data->frame_context_ready = 0;
+#if CONFIG_MULTITHREAD
+    if (pthread_mutex_init(&frame_worker_data->stats_mutex, NULL)) {
+      set_error_detail(ctx, "Failed to allocate frame_worker_data mutex");
+      return VPX_CODEC_MEM_ERROR;
+    }
+
+    if (pthread_cond_init(&frame_worker_data->stats_cond, NULL)) {
+      set_error_detail(ctx, "Failed to allocate frame_worker_data cond");
+      return VPX_CODEC_MEM_ERROR;
+    }
+#endif
+    // If decoding in serial mode, FrameWorker thread could create tile worker
+    // thread or loopfilter thread.
+    frame_worker_data->pbi->max_threads =
+        (ctx->frame_parallel_decode == 0) ? ctx->cfg.threads : 0;
+
+    frame_worker_data->pbi->inv_tile_order = ctx->invert_tile_order;
+    frame_worker_data->pbi->frame_parallel_decode = ctx->frame_parallel_decode;
+    frame_worker_data->pbi->common.frame_parallel_decode =
+        ctx->frame_parallel_decode;
+    worker->hook = (VP9WorkerHook)frame_worker_hook;
+    if (!winterface->reset(worker)) {
+      set_error_detail(ctx, "Frame Worker thread creation failed");
+      return VPX_CODEC_MEM_ERROR;
+    }
+  }
 
   // If postprocessing was enabled by the application and a
   // configuration has not been provided, default it.
@@ -248,20 +414,17 @@ static void init_decoder(vpx_codec_alg_priv_t *ctx) {
     set_default_ppflags(&ctx->postproc_cfg);
 
   init_buffer_callbacks(ctx);
+
+  return VPX_CODEC_OK;
 }
 
 static vpx_codec_err_t decode_one(vpx_codec_alg_priv_t *ctx,
                                   const uint8_t **data, unsigned int data_sz,
                                   void *user_priv, int64_t deadline) {
-  YV12_BUFFER_CONFIG sd;
-  vp9_ppflags_t flags = {0, 0, 0};
-  VP9_COMMON *cm = NULL;
-
+  vp9_ppflags_t flags = {0};
+  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
   (void)deadline;
 
-  vp9_zero(sd);
-  ctx->img_avail = 0;
-
   // Determine the stream parameters. Note that we rely on peek_si to
   // validate that we have a buffer that does not wrap around the top
   // of the heap.
@@ -276,33 +439,73 @@ static vpx_codec_err_t decode_one(vpx_codec_alg_priv_t *ctx,
       return VPX_CODEC_ERROR;
   }
 
-  // Initialize the decoder instance on the first frame
-  if (ctx->pbi == NULL) {
-    init_decoder(ctx);
-    if (ctx->pbi == NULL)
-      return VPX_CODEC_ERROR;
-  }
+  if (!ctx->frame_parallel_decode) {
+    VP9Worker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    frame_worker_data->data = *data;
+    frame_worker_data->data_size = data_sz;
+    frame_worker_data->user_priv = user_priv;
 
     // Set these even if already initialized.  The caller may have changed the
     // decrypt config between frames.
-  ctx->pbi->decrypt_cb = ctx->decrypt_cb;
-  ctx->pbi->decrypt_state = ctx->decrypt_state;
+    frame_worker_data->pbi->decrypt_cb = ctx->decrypt_cb;
+    frame_worker_data->pbi->decrypt_state = ctx->decrypt_state;
 
-  cm = &ctx->pbi->common;
+    worker->had_error = 0;
+    winterface->execute(worker);
 
-  if (vp9_receive_compressed_data(ctx->pbi, data_sz, data))
-    return update_error_state(ctx, &cm->error);
+    // Update data pointer after decode.
+    *data = frame_worker_data->data_end;
+
+    if (worker->had_error)
+      return update_error_state(ctx, &frame_worker_data->pbi->common.error);
+  } else {
+    const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
+    VP9Worker *const worker = &ctx->frame_workers[ctx->next_submit_worker_id];
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    // Copy context from last worker thread to next worker thread.
+    if (ctx->next_submit_worker_id != ctx->last_submit_worker_id)
+      vp9_frameworker_copy_context(
+          &ctx->frame_workers[ctx->next_submit_worker_id],
+          &ctx->frame_workers[ctx->last_submit_worker_id]);
+
+    frame_worker_data->pbi->ready_for_new_data = 0;
+    // Copy the compressed data into worker's internal buffer.
+    // TODO(hkuang): Will all the workers allocate the same size
+    // as the size of the first intra frame be better? This will
+    // avoid too many deallocate and allocate.
+    if (frame_worker_data->scratch_buffer_size < data_sz) {
+      frame_worker_data->scratch_buffer =
+          (uint8_t *)vpx_realloc(frame_worker_data->scratch_buffer, data_sz);
+      if (frame_worker_data->scratch_buffer == NULL) {
+        set_error_detail(ctx, "Failed to reallocate scratch buffer");
+        return VPX_CODEC_MEM_ERROR;
+      }
+      frame_worker_data->scratch_buffer_size = data_sz;
+    }
+    frame_worker_data->data_size = data_sz;
+    vpx_memcpy(frame_worker_data->scratch_buffer, *data, data_sz);
+
+    frame_worker_data->frame_decoded = 0;
+    frame_worker_data->frame_context_ready = 0;
+    frame_worker_data->data = frame_worker_data->scratch_buffer;
+    frame_worker_data->user_priv = user_priv;
+
+    if (ctx->next_submit_worker_id != ctx->last_submit_worker_id)
+      ctx->last_submit_worker_id =
+          (ctx->last_submit_worker_id + 1) % ctx->num_frame_workers;
+
+    ctx->next_submit_worker_id =
+        (ctx->next_submit_worker_id + 1) % ctx->num_frame_workers;
+
+    --ctx->available_threads;
+    worker->had_error = 0;
+    winterface->launch(worker);
+  }
 
   if (ctx->base.init_flags & VPX_CODEC_USE_POSTPROC)
     set_ppflags(ctx, &flags);
 
-  if (vp9_get_raw_frame(ctx->pbi, &sd, &flags))
-    return update_error_state(ctx, &cm->error);
-
-  yuvconfig2image(&ctx->img, &sd, user_priv);
-  ctx->img.fb_priv = cm->frame_bufs[cm->new_fb_idx].raw_frame_buffer.priv;
-  ctx->img_avail = 1;
-
   return VPX_CODEC_OK;
 }
 
@@ -381,6 +584,30 @@ static vpx_codec_err_t parse_superframe_index(const uint8_t *data,
   return VPX_CODEC_OK;
 }
 
+static void wait_worker_and_cache_frame(vpx_codec_alg_priv_t *ctx) {
+  YV12_BUFFER_CONFIG sd;
+  vp9_ppflags_t flags = {0, 0, 0};
+  const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
+  VP9Worker *const worker = &ctx->frame_workers[ctx->next_output_worker_id];
+  FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+  ctx->next_output_worker_id =
+      (ctx->next_output_worker_id + 1) % ctx->num_frame_workers;
+  winterface->sync(worker);
+  ++ctx->available_threads;
+  if (vp9_get_raw_frame(frame_worker_data->pbi, &sd, &flags) == 0) {
+    VP9_COMMON *const cm = &frame_worker_data->pbi->common;
+    RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+    ctx->frame_cache[ctx->frame_cache_write].fb_idx = cm->new_fb_idx;
+    yuvconfig2image(&ctx->frame_cache[ctx->frame_cache_write].img, &sd,
+                    frame_worker_data->user_priv);
+    ctx->frame_cache[ctx->frame_cache_write].img.fb_priv =
+        frame_bufs[cm->new_fb_idx].raw_frame_buffer.priv;
+    ctx->frame_cache_write =
+        (ctx->frame_cache_write + 1) % FRAME_CACHE_SIZE;
+    ++ctx->num_cache_frames;
+  }
+}
+
 static vpx_codec_err_t decoder_decode(vpx_codec_alg_priv_t *ctx,
                                       const uint8_t *data, unsigned int data_sz,
                                       void *user_priv, long deadline) {
@@ -390,8 +617,20 @@ static vpx_codec_err_t decoder_decode(vpx_codec_alg_priv_t *ctx,
   uint32_t frame_sizes[8];
   int frame_count;
 
-  if (data == NULL || data_sz == 0)
-    return VPX_CODEC_INVALID_PARAM;
+  if (data == NULL && data_sz == 0) {
+    ctx->flushed = 1;
+    return VPX_CODEC_OK;
+  }
+
+  // Reset flushed when receiving a valid frame.
+  ctx->flushed = 0;
+
+  // Initialize the decoder workers on the first frame.
+  if (ctx->frame_workers == NULL) {
+    const vpx_codec_err_t res = init_decoder(ctx);
+    if (res != VPX_CODEC_OK)
+      return res;
+  }
 
   res = parse_superframe_index(data, data_sz, frame_sizes, &frame_count,
                                ctx->decrypt_cb, ctx->decrypt_state);
@@ -409,30 +648,46 @@ static vpx_codec_err_t decoder_decode(vpx_codec_alg_priv_t *ctx,
       for (i = 0; i < frame_count; ++i) {
         const uint8_t *data_start_copy = data_start;
         const uint32_t frame_size = frame_sizes[i];
-        vpx_codec_err_t res;
         if (data_start < data
             || frame_size > (uint32_t) (data_end - data_start)) {
-          ctx->base.err_detail = "Invalid frame size in index";
+          set_error_detail(ctx, "Invalid frame size in index");
           return VPX_CODEC_CORRUPT_FRAME;
         }
 
+        if (ctx->available_threads == 0) {
+          // No more threads for decoding. Wait until the next output worker
+          // finishes decoding. Then copy the decoded frame into cache.
+          if (ctx->num_cache_frames < FRAME_CACHE_SIZE) {
+            wait_worker_and_cache_frame(ctx);
+          } else {
+            // TODO(hkuang): Add unit test to test this path.
+            set_error_detail(ctx, "Frame output cache is full.");
+            return VPX_CODEC_ERROR;
+          }
+        }
+
         res = decode_one(ctx, &data_start_copy, frame_size, user_priv,
                          deadline);
         if (res != VPX_CODEC_OK)
           return res;
-
         data_start += frame_size;
       }
     } else {
-      res = decode_one(ctx, &data_start, data_sz, user_priv, deadline);
+      if (ctx->available_threads == 0) {
+        // No more threads for decoding. Wait until the next output worker
+        // finishes decoding. Then copy the decoded frame into cache.
+        if (ctx->num_cache_frames < FRAME_CACHE_SIZE) {
+          wait_worker_and_cache_frame(ctx);
+        } else {
+          // TODO(hkuang): Add unit test to test this path.
+          set_error_detail(ctx, "Frame output cache is full.");
+          return VPX_CODEC_ERROR;
+        }
+      }
+
+      res = decode_one(ctx, &data, data_sz, user_priv, deadline);
       if (res != VPX_CODEC_OK)
         return res;
-
-      // Extra data detected after the frame.
-      if (data_start < data_end - 1) {
-        ctx->base.err_detail = "Fail to decode frame in parallel mode";
-        return VPX_CODEC_INCAPABLE;
-      }
     }
   } else {
     // Decode in serial mode.
@@ -445,7 +700,7 @@ static vpx_codec_err_t decoder_decode(vpx_codec_alg_priv_t *ctx,
         vpx_codec_err_t res;
         if (data_start < data
             || frame_size > (uint32_t) (data_end - data_start)) {
-          ctx->base.err_detail = "Invalid frame size in index";
+          set_error_detail(ctx, "Invalid frame size in index");
           return VPX_CODEC_CORRUPT_FRAME;
         }
 
@@ -476,23 +731,73 @@ static vpx_codec_err_t decoder_decode(vpx_codec_alg_priv_t *ctx,
     }
   }
 
-  return VPX_CODEC_OK;
+  return res;
+}
+
+static void release_last_output_frame(vpx_codec_alg_priv_t *ctx) {
+  RefCntBuffer *const frame_bufs = ctx->buffer_pool->frame_bufs;
+  // Decrease reference count of last output frame in frame parallel mode.
+  if (ctx->frame_parallel_decode && ctx->last_show_frame >= 0) {
+    BufferPool *const pool = ctx->buffer_pool;
+    lock_buffer_pool(pool);
+    decrease_ref_count(ctx->last_show_frame, frame_bufs, pool);
+    unlock_buffer_pool(pool);
+  }
 }
 
 static vpx_image_t *decoder_get_frame(vpx_codec_alg_priv_t *ctx,
                                       vpx_codec_iter_t *iter) {
   vpx_image_t *img = NULL;
 
-  if (ctx->img_avail) {
+  // Only return frame when all the cpu are busy or
+  // application fluhsed the decoder in frame parallel decode.
+  if (ctx->frame_parallel_decode && ctx->available_threads > 0 &&
+      !ctx->flushed) {
+    return img;
+  }
+
+  // Output the frames in the cache first.
+  if (ctx->num_cache_frames > 0) {
+    release_last_output_frame(ctx);
+    ctx->last_show_frame  = ctx->frame_cache[ctx->frame_cache_read].fb_idx;
+    img = &ctx->frame_cache[ctx->frame_cache_read].img;
+    ctx->frame_cache_read = (ctx->frame_cache_read + 1) % FRAME_CACHE_SIZE;
+    --ctx->num_cache_frames;
+    return img;
+  }
+
   // iter acts as a flip flop, so an image is only returned on the first
   // call to get_frame.
-    if (!(*iter)) {
+  if (*iter == NULL && ctx->frame_workers != NULL) {
+    do {
+      YV12_BUFFER_CONFIG sd;
+      vp9_ppflags_t flags = {0, 0, 0};
+      const VP9WorkerInterface *const winterface = vp9_get_worker_interface();
+      VP9Worker *const worker =
+          &ctx->frame_workers[ctx->next_output_worker_id];
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      ctx->next_output_worker_id =
+          (ctx->next_output_worker_id + 1) % ctx->num_frame_workers;
+      // Wait for the frame from worker thread.
+      if (!winterface->sync(worker)) {
+        // Decoding failed. Release the worker thread.
+        ++ctx->available_threads;
+        if (ctx->flushed != 1)
+          return img;
+      } else if (vp9_get_raw_frame(frame_worker_data->pbi, &sd, &flags) == 0) {
+        VP9_COMMON *const cm = &frame_worker_data->pbi->common;
+        RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;
+        ++ctx->available_threads;
+        release_last_output_frame(ctx);
+        ctx->last_show_frame = frame_worker_data->pbi->common.new_fb_idx;
+        yuvconfig2image(&ctx->img, &sd, frame_worker_data->user_priv);
+        ctx->img.fb_priv = frame_bufs[cm->new_fb_idx].raw_frame_buffer.priv;
         img = &ctx->img;
-      *iter = img;
+        return img;
       }
+    } while (ctx->next_output_worker_id != ctx->next_submit_worker_id);
   }
-  ctx->img_avail = 0;
-
   return img;
 }
 
@@ -502,7 +807,7 @@ static vpx_codec_err_t decoder_set_fb_fn(
     vpx_release_frame_buffer_cb_fn_t cb_release, void *cb_priv) {
   if (cb_get == NULL || cb_release == NULL) {
     return VPX_CODEC_INVALID_PARAM;
-  } else if (ctx->pbi == NULL) {
+  } else if (ctx->frame_workers == NULL) {
     // If the decoder has already been initialized, do not accept changes to
     // the frame buffer functions.
     ctx->get_ext_fb_cb = cb_get;
@@ -518,12 +823,19 @@ static vpx_codec_err_t ctrl_set_reference(vpx_codec_alg_priv_t *ctx,
                                           va_list args) {
   vpx_ref_frame_t *const data = va_arg(args, vpx_ref_frame_t *);
 
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return VPX_CODEC_INCAPABLE;
+  }
+
   if (data) {
     vpx_ref_frame_t *const frame = (vpx_ref_frame_t *)data;
     YV12_BUFFER_CONFIG sd;
-
+    VP9Worker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
     image2yuvconfig(&frame->img, &sd);
-    return vp9_set_reference_dec(&ctx->pbi->common,
+    return vp9_set_reference_dec(&frame_worker_data->pbi->common,
                                  (VP9_REFFRAME)frame->frame_type, &sd);
   } else {
     return VPX_CODEC_INVALID_PARAM;
@@ -534,13 +846,19 @@ static vpx_codec_err_t ctrl_copy_reference(vpx_codec_alg_priv_t *ctx,
                                            va_list args) {
   vpx_ref_frame_t *data = va_arg(args, vpx_ref_frame_t *);
 
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return VPX_CODEC_INCAPABLE;
+  }
+
   if (data) {
-    vpx_ref_frame_t *frame = (vpx_ref_frame_t *)data;
+    vpx_ref_frame_t *frame = (vpx_ref_frame_t *) data;
     YV12_BUFFER_CONFIG sd;
-
+    VP9Worker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
     image2yuvconfig(&frame->img, &sd);
-
-    return vp9_copy_reference_dec(ctx->pbi,
+    return vp9_copy_reference_dec(frame_worker_data->pbi,
                                   (VP9_REFFRAME)frame->frame_type, &sd);
   } else {
     return VPX_CODEC_INVALID_PARAM;
@@ -551,10 +869,17 @@ static vpx_codec_err_t ctrl_get_reference(vpx_codec_alg_priv_t *ctx,
                                           va_list args) {
   vp9_ref_frame_t *data = va_arg(args, vp9_ref_frame_t *);
 
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return VPX_CODEC_INCAPABLE;
+  }
+
   if (data) {
     YV12_BUFFER_CONFIG* fb;
-
-    vp9_get_reference_dec(ctx->pbi, data->idx, &fb);
+    VP9Worker *const worker = ctx->frame_workers;
+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;
+    vp9_get_reference_dec(frame_worker_data->pbi, data->idx, &fb);
     yuvconfig2image(&data->img, fb, NULL);
     return VPX_CODEC_OK;
   } else {
@@ -592,27 +917,42 @@ static vpx_codec_err_t ctrl_get_last_ref_updates(vpx_codec_alg_priv_t *ctx,
                                                  va_list args) {
   int *const update_info = va_arg(args, int *);
 
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return VPX_CODEC_INCAPABLE;
+  }
+
   if (update_info) {
-    if (ctx->pbi)
-      *update_info = ctx->pbi->refresh_frame_flags;
-    else
+    if (ctx->frame_workers) {
+      VP9Worker *const worker = ctx->frame_workers;
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      *update_info = frame_worker_data->pbi->refresh_frame_flags;
+    } else {
       return VPX_CODEC_ERROR;
+    }
     return VPX_CODEC_OK;
   } else {
     return VPX_CODEC_INVALID_PARAM;
   }
 }
 
-
 static vpx_codec_err_t ctrl_get_frame_corrupted(vpx_codec_alg_priv_t *ctx,
                                                 va_list args) {
   int *corrupted = va_arg(args, int *);
 
   if (corrupted) {
-    if (ctx->pbi)
-      *corrupted = ctx->pbi->common.frame_to_show->corrupted;
-    else
+    if (ctx->frame_workers) {
+      VP9Worker *const worker = ctx->frame_workers;
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      RefCntBuffer *const frame_bufs =
+          frame_worker_data->pbi->common.buffer_pool->frame_bufs;
+      *corrupted = frame_bufs[ctx->last_show_frame].buf.corrupted;
+    } else {
       return VPX_CODEC_ERROR;
+    }
     return VPX_CODEC_OK;
   } else {
     return VPX_CODEC_INVALID_PARAM;
@@ -623,9 +963,18 @@ static vpx_codec_err_t ctrl_get_display_size(vpx_codec_alg_priv_t *ctx,
                                              va_list args) {
   int *const display_size = va_arg(args, int *);
 
+  // Only support this function in serial decode.
+  if (ctx->frame_parallel_decode) {
+    set_error_detail(ctx, "Not supported in frame parallel decode");
+    return VPX_CODEC_INCAPABLE;
+  }
+
   if (display_size) {
-    if (ctx->pbi) {
-      const VP9_COMMON *const cm = &ctx->pbi->common;
+    if (ctx->frame_workers) {
+      VP9Worker *const worker = ctx->frame_workers;
+      FrameWorkerData *const frame_worker_data =
+          (FrameWorkerData *)worker->data1;
+      const VP9_COMMON *const cm = &frame_worker_data->pbi->common;
       display_size[0] = cm->display_width;
       display_size[1] = cm->display_height;
     } else {

vp9/vp9dx.mk

@@ -31,8 +31,6 @@ VP9_DX_SRCS-yes += decoder/vp9_decodemv.h
 VP9_DX_SRCS-yes += decoder/vp9_detokenize.h
 VP9_DX_SRCS-yes += decoder/vp9_decoder.c
 VP9_DX_SRCS-yes += decoder/vp9_decoder.h
-VP9_DX_SRCS-yes += decoder/vp9_thread.c
-VP9_DX_SRCS-yes += decoder/vp9_thread.h
 VP9_DX_SRCS-yes += decoder/vp9_dsubexp.c
 VP9_DX_SRCS-yes += decoder/vp9_dsubexp.h
 

vpx/vpx_frame_buffer.h

@@ -22,8 +22,11 @@ extern "C" {
 #include "./vpx_integer.h"
 
 /*!\brief The maximum number of work buffers used by libvpx.
+ *  Support maximum 4 threads to decode video in parallel.
+ *  Each thread will use one work buffer.
+ * TODO(hkuang): Add support to set number of worker threads dynamically.
  */
-#define VPX_MAXIMUM_WORK_BUFFERS 1
+#define VPX_MAXIMUM_WORK_BUFFERS 8
 
 /*!\brief The maximum number of reference buffers that a VP9 encoder may use.
  */

webmdec.cc

@@ -41,6 +41,7 @@ void reset(struct WebmInputContext *const webm_ctx) {
   webm_ctx->block_frame_index = 0;
   webm_ctx->video_track_index = 0;
   webm_ctx->timestamp_ns = 0;
+  webm_ctx->is_key_frame = false;
 }
 
 void get_first_cluster(struct WebmInputContext *const webm_ctx) {
@@ -182,6 +183,7 @@ int webm_read_frame(struct WebmInputContext *webm_ctx,
   }
   *bytes_in_buffer = frame.len;
   webm_ctx->timestamp_ns = block->GetTime(cluster);
+  webm_ctx->is_key_frame = block->IsKey();
 
   mkvparser::MkvReader *const reader =
       reinterpret_cast<mkvparser::MkvReader*>(webm_ctx->reader);

webmdec.h

@@ -28,6 +28,7 @@ struct WebmInputContext {
   int block_frame_index;
   int video_track_index;
   uint64_t timestamp_ns;
+  int is_key_frame;
 };
 
 // Checks if the input is a WebM file. If so, initializes WebMInputContext so