Bug 1380118 - Update aom library. draft
authorRalph Giles <giles@mozilla.com>
Mon, 10 Jul 2017 15:33:39 -0700
changeset 619853 5055b72960bbae413061b4712f6d00befde2b266
parent 619852 8fa2880513e28af0d842da8515a1d29d6593742a
child 640526 2350cef3c7df5835e418c639c5769e0b3ec02abb
push id71850
push userbmo:giles@thaumas.net
push dateWed, 02 Aug 2017 18:38:05 +0000
bugs1380118
milestone56.0a1
Bug 1380118 - Update aom library. Update to git commit f5bdeac22930ff4c6b219be49c843db35970b918 to pick up recent changes. MozReview-Commit-ID: E9CahggMR3
dom/media/platforms/agnostic/AOMDecoder.cpp
media/libaom/README_MOZILLA
media/libaom/config/generic/aom_config.asm
media/libaom/config/generic/aom_config.h
media/libaom/config/generic/aom_dsp_rtcd.h
media/libaom/config/generic/av1_rtcd.h
media/libaom/config/linux/arm/aom_config.asm
media/libaom/config/linux/arm/aom_config.h
media/libaom/config/linux/arm/aom_dsp_rtcd.h
media/libaom/config/linux/arm/av1_rtcd.h
media/libaom/config/linux/ia32/aom_config.asm
media/libaom/config/linux/ia32/aom_config.h
media/libaom/config/linux/ia32/aom_dsp_rtcd.h
media/libaom/config/linux/ia32/av1_rtcd.h
media/libaom/config/linux/x64/aom_config.asm
media/libaom/config/linux/x64/aom_config.h
media/libaom/config/linux/x64/aom_dsp_rtcd.h
media/libaom/config/linux/x64/av1_rtcd.h
media/libaom/config/mac/x64/aom_config.asm
media/libaom/config/mac/x64/aom_config.h
media/libaom/config/mac/x64/aom_dsp_rtcd.h
media/libaom/config/mac/x64/av1_rtcd.h
media/libaom/config/win/ia32/aom_config.asm
media/libaom/config/win/ia32/aom_config.h
media/libaom/config/win/ia32/aom_dsp_rtcd.h
media/libaom/config/win/ia32/av1_rtcd.h
media/libaom/config/win/mingw32/aom_config.asm
media/libaom/config/win/mingw32/aom_config.c
media/libaom/config/win/mingw32/aom_config.h
media/libaom/config/win/mingw32/aom_dsp_rtcd.h
media/libaom/config/win/mingw32/aom_scale_rtcd.h
media/libaom/config/win/mingw32/av1_rtcd.h
media/libaom/config/win/x64/aom_config.asm
media/libaom/config/win/x64/aom_config.h
media/libaom/config/win/x64/aom_dsp_rtcd.h
media/libaom/config/win/x64/av1_rtcd.h
media/libaom/moz.build
media/libaom/sources.mozbuild
python/mozbuild/mozbuild/vendor_aom.py
third_party/aom/CMakeLists.txt
third_party/aom/README
third_party/aom/README.md
third_party/aom/aom/aom_decoder.h
third_party/aom/aom/aom_encoder.h
third_party/aom/aom/aom_image.h
third_party/aom/aom/aom_integer.h
third_party/aom/aom/aomcx.h
third_party/aom/aom_dsp/aom_convolve.c
third_party/aom/aom_dsp/aom_dsp.cmake
third_party/aom/aom_dsp/aom_dsp.mk
third_party/aom/aom_dsp/aom_dsp_common.h
third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl
third_party/aom/aom_dsp/aom_filter.h
third_party/aom/aom_dsp/bitreader.h
third_party/aom/aom_dsp/bitwriter.h
third_party/aom/aom_dsp/blend_a64_mask.c
third_party/aom/aom_dsp/fwd_txfm.c
third_party/aom/aom_dsp/fwd_txfm.h
third_party/aom/aom_dsp/intrapred.c
third_party/aom/aom_dsp/inv_txfm.c
third_party/aom/aom_dsp/inv_txfm.h
third_party/aom/aom_dsp/mips/fwd_dct32x32_msa.c
third_party/aom/aom_dsp/mips/fwd_txfm_msa.c
third_party/aom/aom_dsp/prob.c
third_party/aom/aom_dsp/prob.h
third_party/aom/aom_dsp/quantize.c
third_party/aom/aom_dsp/sad.c
third_party/aom/aom_dsp/simd/v256_intrinsics_v128.h
third_party/aom/aom_dsp/txfm_common.h
third_party/aom/aom_dsp/variance.c
third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c
third_party/aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c
third_party/aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c
third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm
third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm
third_party/aom/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm
third_party/aom/aom_dsp/x86/highbd_variance_sse2.c
third_party/aom/aom_dsp/x86/inv_txfm_sse2.c
third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c
third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c
third_party/aom/aom_dsp/x86/obmc_sad_sse4.c
third_party/aom/aom_dsp/x86/obmc_variance_sse4.c
third_party/aom/aom_dsp/x86/sad4d_sse2.asm
third_party/aom/aom_dsp/x86/sad_highbd_avx2.c
third_party/aom/aom_dsp/x86/sad_sse2.asm
third_party/aom/aom_dsp/x86/variance_sse2.c
third_party/aom/aom_mem/aom_mem.cmake
third_party/aom/aom_ports/aom_ports.cmake
third_party/aom/aom_scale/aom_scale.cmake
third_party/aom/aom_scale/yv12config.h
third_party/aom/aom_util/aom_util.cmake
third_party/aom/aomdec.c
third_party/aom/aomenc.c
third_party/aom/av1/av1.cmake
third_party/aom/av1/av1_common.mk
third_party/aom/av1/av1_cx.mk
third_party/aom/av1/av1_cx_iface.c
third_party/aom/av1/av1_dx_iface.c
third_party/aom/av1/av1_iface_common.h
third_party/aom/av1/common/alloccommon.c
third_party/aom/av1/common/arm/neon/iht4x4_add_neon.c
third_party/aom/av1/common/arm/neon/iht8x8_add_neon.c
third_party/aom/av1/common/av1_fwd_txfm1d_cfg.h
third_party/aom/av1/common/av1_fwd_txfm2d.c
third_party/aom/av1/common/av1_inv_txfm1d_cfg.h
third_party/aom/av1/common/av1_inv_txfm2d.c
third_party/aom/av1/common/av1_loopfilter.c
third_party/aom/av1/common/av1_loopfilter.h
third_party/aom/av1/common/av1_rtcd_defs.pl
third_party/aom/av1/common/av1_txfm.h
third_party/aom/av1/common/blockd.c
third_party/aom/av1/common/blockd.h
third_party/aom/av1/common/cdef.c
third_party/aom/av1/common/cdef.h
third_party/aom/av1/common/cfl.c
third_party/aom/av1/common/cfl.h
third_party/aom/av1/common/clpf.c
third_party/aom/av1/common/clpf.h
third_party/aom/av1/common/common_data.h
third_party/aom/av1/common/convolve.c
third_party/aom/av1/common/convolve.h
third_party/aom/av1/common/daala_tx.c
third_party/aom/av1/common/daala_tx.h
third_party/aom/av1/common/entropy.c
third_party/aom/av1/common/entropy.h
third_party/aom/av1/common/entropymode.c
third_party/aom/av1/common/entropymode.h
third_party/aom/av1/common/entropymv.c
third_party/aom/av1/common/entropymv.h
third_party/aom/av1/common/enums.h
third_party/aom/av1/common/filter.h
third_party/aom/av1/common/generic_code.c
third_party/aom/av1/common/generic_code.h
third_party/aom/av1/common/idct.c
third_party/aom/av1/common/idct.h
third_party/aom/av1/common/mips/dspr2/av1_itrans16_dspr2.c
third_party/aom/av1/common/mips/dspr2/av1_itrans4_dspr2.c
third_party/aom/av1/common/mips/dspr2/av1_itrans8_dspr2.c
third_party/aom/av1/common/mips/msa/av1_idct16x16_msa.c
third_party/aom/av1/common/mips/msa/av1_idct4x4_msa.c
third_party/aom/av1/common/mips/msa/av1_idct8x8_msa.c
third_party/aom/av1/common/mv.h
third_party/aom/av1/common/mvref_common.c
third_party/aom/av1/common/mvref_common.h
third_party/aom/av1/common/od_dering.c
third_party/aom/av1/common/od_dering.h
third_party/aom/av1/common/odintrin.h
third_party/aom/av1/common/onyxc_int.h
third_party/aom/av1/common/pred_common.c
third_party/aom/av1/common/pred_common.h
third_party/aom/av1/common/quant_common.c
third_party/aom/av1/common/reconinter.c
third_party/aom/av1/common/reconinter.h
third_party/aom/av1/common/reconintra.c
third_party/aom/av1/common/reconintra.h
third_party/aom/av1/common/resize.c
third_party/aom/av1/common/resize.h
third_party/aom/av1/common/restoration.c
third_party/aom/av1/common/scale.c
third_party/aom/av1/common/scale.h
third_party/aom/av1/common/scan.c
third_party/aom/av1/common/scan.h
third_party/aom/av1/common/seg_common.h
third_party/aom/av1/common/tile_common.c
third_party/aom/av1/common/tile_common.h
third_party/aom/av1/common/txb_common.c
third_party/aom/av1/common/txb_common.h
third_party/aom/av1/common/warped_motion.c
third_party/aom/av1/common/warped_motion.h
third_party/aom/av1/common/x86/av1_convolve_ssse3.c
third_party/aom/av1/common/x86/av1_fwd_txfm2d_sse4.c
third_party/aom/av1/common/x86/av1_txfm1d_sse4.h
third_party/aom/av1/common/x86/convolve_2d_sse2.c
third_party/aom/av1/common/x86/convolve_avx2.c
third_party/aom/av1/common/x86/highbd_convolve_2d_ssse3.c
third_party/aom/av1/common/x86/highbd_warp_plane_ssse3.c
third_party/aom/av1/common/x86/hybrid_inv_txfm_avx2.c
third_party/aom/av1/common/x86/idct_intrin_sse2.c
third_party/aom/av1/common/x86/selfguided_sse4.c
third_party/aom/av1/common/x86/warp_plane_sse2.c
third_party/aom/av1/common/x86/warp_plane_ssse3.c
third_party/aom/av1/decoder/decodeframe.c
third_party/aom/av1/decoder/decodemv.c
third_party/aom/av1/decoder/decodemv.h
third_party/aom/av1/decoder/decoder.c
third_party/aom/av1/decoder/decoder.h
third_party/aom/av1/decoder/decodetxb.c
third_party/aom/av1/decoder/decodetxb.h
third_party/aom/av1/decoder/detokenize.c
third_party/aom/av1/decoder/detokenize.h
third_party/aom/av1/decoder/inspection.c
third_party/aom/av1/encoder/aq_cyclicrefresh.c
third_party/aom/av1/encoder/av1_quantize.c
third_party/aom/av1/encoder/av1_quantize.h
third_party/aom/av1/encoder/bgsprite.c
third_party/aom/av1/encoder/bgsprite.h
third_party/aom/av1/encoder/bitstream.c
third_party/aom/av1/encoder/bitstream.h
third_party/aom/av1/encoder/block.h
third_party/aom/av1/encoder/context_tree.c
third_party/aom/av1/encoder/context_tree.h
third_party/aom/av1/encoder/cost.c
third_party/aom/av1/encoder/cost.h
third_party/aom/av1/encoder/dct.c
third_party/aom/av1/encoder/encodeframe.c
third_party/aom/av1/encoder/encodeframe.h
third_party/aom/av1/encoder/encodemb.c
third_party/aom/av1/encoder/encodemb.h
third_party/aom/av1/encoder/encodemv.c
third_party/aom/av1/encoder/encodemv.h
third_party/aom/av1/encoder/encoder.c
third_party/aom/av1/encoder/encoder.h
third_party/aom/av1/encoder/encodetxb.c
third_party/aom/av1/encoder/encodetxb.h
third_party/aom/av1/encoder/ethread.c
third_party/aom/av1/encoder/firstpass.c
third_party/aom/av1/encoder/firstpass.h
third_party/aom/av1/encoder/global_motion.c
third_party/aom/av1/encoder/global_motion.h
third_party/aom/av1/encoder/hybrid_fwd_txfm.c
third_party/aom/av1/encoder/hybrid_fwd_txfm.h
third_party/aom/av1/encoder/mcomp.c
third_party/aom/av1/encoder/mcomp.h
third_party/aom/av1/encoder/palette.c
third_party/aom/av1/encoder/palette.h
third_party/aom/av1/encoder/pickcdef.c
third_party/aom/av1/encoder/picklpf.c
third_party/aom/av1/encoder/picklpf.h
third_party/aom/av1/encoder/pickrst.c
third_party/aom/av1/encoder/ransac.c
third_party/aom/av1/encoder/ratectrl.c
third_party/aom/av1/encoder/ratectrl.h
third_party/aom/av1/encoder/rd.c
third_party/aom/av1/encoder/rd.h
third_party/aom/av1/encoder/rdopt.c
third_party/aom/av1/encoder/rdopt.h
third_party/aom/av1/encoder/segmentation.c
third_party/aom/av1/encoder/speed_features.c
third_party/aom/av1/encoder/speed_features.h
third_party/aom/av1/encoder/temporal_filter.c
third_party/aom/av1/encoder/temporal_filter.h
third_party/aom/av1/encoder/tokenize.c
third_party/aom/av1/encoder/tokenize.h
third_party/aom/av1/encoder/x86/av1_highbd_quantize_avx2.c
third_party/aom/av1/encoder/x86/av1_highbd_quantize_sse4.c
third_party/aom/av1/encoder/x86/av1_quantize_avx2.c
third_party/aom/av1/encoder/x86/dct_intrin_sse2.c
third_party/aom/av1/encoder/x86/error_intrin_avx2.c
third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c
third_party/aom/av1/encoder/x86/hybrid_fwd_txfm_avx2.c
third_party/aom/build/cmake/aom_config_defaults.cmake
third_party/aom/build/cmake/aom_configure.cmake
third_party/aom/build/cmake/aom_optimization.cmake
third_party/aom/build/cmake/compiler_flags.cmake
third_party/aom/build/cmake/exports.cmake
third_party/aom/build/cmake/exports_sources.cmake
third_party/aom/build/cmake/generate_exports.cmake
third_party/aom/build/cmake/toolchains/arm-ios-common.cmake
third_party/aom/build/cmake/toolchains/arm64-linux-gcc.cmake
third_party/aom/build/cmake/toolchains/armv7-ios.cmake
third_party/aom/build/cmake/toolchains/armv7-linux-gcc.cmake
third_party/aom/build/cmake/toolchains/armv7s-ios.cmake
third_party/aom/build/cmake/toolchains/mips32-linux-gcc.cmake
third_party/aom/build/cmake/toolchains/mips64-linux-gcc.cmake
third_party/aom/build/cmake/toolchains/x86-ios-simulator.cmake
third_party/aom/build/cmake/toolchains/x86-macos.cmake
third_party/aom/build/cmake/util.cmake
third_party/aom/build/make/gen_msvs_sln.sh
third_party/aom/build/make/gen_msvs_vcxproj.sh
third_party/aom/configure
third_party/aom/docs.cmake
third_party/aom/examples/analyzer.cc
third_party/aom/examples/inspect.c
third_party/aom/libs.mk
third_party/aom/test/active_map_refresh_test.cc
third_party/aom/test/active_map_test.cc
third_party/aom/test/altref_test.cc
third_party/aom/test/ans_codec_test.cc
third_party/aom/test/aq_segment_test.cc
third_party/aom/test/arf_freq_test.cc
third_party/aom/test/av1_convolve_2d_test.cc
third_party/aom/test/av1_convolve_2d_test_util.cc
third_party/aom/test/av1_convolve_2d_test_util.h
third_party/aom/test/av1_convolve_optimz_test.cc
third_party/aom/test/av1_convolve_test.cc
third_party/aom/test/av1_dct_test.cc
third_party/aom/test/av1_ext_tile_test.cc
third_party/aom/test/av1_fht16x16_test.cc
third_party/aom/test/av1_fht16x32_test.cc
third_party/aom/test/av1_fht16x8_test.cc
third_party/aom/test/av1_fht32x16_test.cc
third_party/aom/test/av1_fht32x32_test.cc
third_party/aom/test/av1_fht4x4_test.cc
third_party/aom/test/av1_fht4x8_test.cc
third_party/aom/test/av1_fht64x64_test.cc
third_party/aom/test/av1_fht8x16_test.cc
third_party/aom/test/av1_fht8x4_test.cc
third_party/aom/test/av1_fht8x8_test.cc
third_party/aom/test/av1_fwd_txfm1d_test.cc
third_party/aom/test/av1_inv_txfm1d_test.cc
third_party/aom/test/av1_inv_txfm2d_test.cc
third_party/aom/test/av1_inv_txfm_test.cc
third_party/aom/test/block_error_test.cc
third_party/aom/test/borders_test.cc
third_party/aom/test/coding_path_sync.cc
third_party/aom/test/convolve_round_test.cc
third_party/aom/test/cpu_speed_test.cc
third_party/aom/test/datarate_test.cc
third_party/aom/test/dct16x16_test.cc
third_party/aom/test/dct32x32_test.cc
third_party/aom/test/decode_api_test.cc
third_party/aom/test/decode_perf_test.cc
third_party/aom/test/encode_api_test.cc
third_party/aom/test/encode_perf_test.cc
third_party/aom/test/encode_test_driver.cc
third_party/aom/test/encode_test_driver.h
third_party/aom/test/encoder_parms_get_to_decoder.cc
third_party/aom/test/end_to_end_test.cc
third_party/aom/test/error_resilience_test.cc
third_party/aom/test/ethread_test.cc
third_party/aom/test/fdct4x4_test.cc
third_party/aom/test/fdct8x8_test.cc
third_party/aom/test/frame_size_tests.cc
third_party/aom/test/hiprec_convolve_test_util.cc
third_party/aom/test/idct_test.cc
third_party/aom/test/level_test.cc
third_party/aom/test/lossless_test.cc
third_party/aom/test/lpf_8_test.cc
third_party/aom/test/masked_sad_test.cc
third_party/aom/test/masked_variance_test.cc
third_party/aom/test/motion_vector_test.cc
third_party/aom/test/partial_idct_test.cc
third_party/aom/test/quantize_func_test.cc
third_party/aom/test/quantize_test.cc
third_party/aom/test/realtime_test.cc
third_party/aom/test/resize_test.cc
third_party/aom/test/sad_test.cc
third_party/aom/test/selfguided_filter_test.cc
third_party/aom/test/simd_cmp_impl.h
third_party/aom/test/subtract_test.cc
third_party/aom/test/superframe_test.cc
third_party/aom/test/test.cmake
third_party/aom/test/test.mk
third_party/aom/test/test_data_download_worker.cmake
third_party/aom/test/test_runner.cmake
third_party/aom/test/tile_independence_test.cc
third_party/aom/test/tools_common.sh
third_party/aom/test/transform_test_base.h
third_party/aom/test/user_priv_test.cc
third_party/aom/test/util.h
third_party/aom/test/variance_test.cc
third_party/aom/test/warp_filter_test.cc
third_party/aom/test/warp_filter_test_util.cc
third_party/aom/test/warp_filter_test_util.h
third_party/aom/third_party/libwebm/README.libaom
third_party/aom/third_party/libwebm/common/file_util.cc
third_party/aom/third_party/libwebm/common/file_util.h
third_party/aom/third_party/libwebm/common/hdr_util.cc
third_party/aom/third_party/libwebm/common/hdr_util.h
third_party/aom/third_party/libwebm/common/webmids.h
third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxer.cc
third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxer.h
third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxerutil.cc
third_party/aom/third_party/libwebm/mkvmuxer/mkvmuxerutil.h
third_party/aom/third_party/libwebm/mkvmuxer/mkvwriter.cc
third_party/aom/third_party/libwebm/mkvparser/mkvparser.cc
third_party/aom/third_party/libwebm/mkvparser/mkvparser.h
third_party/aom/third_party/libwebm/mkvparser/mkvreader.cc
third_party/aom/tools.mk
third_party/aom/tools/aggregate_entropy_stats.py
third_party/aom/tools/aom_entropy_optimizer.c
third_party/aom/tools/build_inspector.sh
--- a/dom/media/platforms/agnostic/AOMDecoder.cpp
+++ b/dom/media/platforms/agnostic/AOMDecoder.cpp
@@ -222,17 +222,17 @@ AOMDecoder::IsAV1(const nsACString& aMim
 /* static */
 bool
 AOMDecoder::IsSupportedCodec(const nsAString& aCodecType)
 {
   // While AV1 is under development, we describe support
   // for a specific aom commit hash so sites can check
   // compatibility.
   auto version = NS_ConvertASCIItoUTF16("av1.experimental.");
-  version.AppendLiteral("aadbb0251996c8ebb8310567bea330ab7ae9abe4");
+  version.AppendLiteral("f5bdeac22930ff4c6b219be49c843db35970b918");
   return aCodecType.EqualsLiteral("av1") ||
          aCodecType.Equals(version);
 }
 
 /* static */
 bool
 AOMDecoder::IsKeyframe(Span<const uint8_t> aBuffer) {
   aom_codec_stream_info_t info;
--- a/media/libaom/README_MOZILLA
+++ b/media/libaom/README_MOZILLA
@@ -13,9 +13,9 @@ To update the library source and build c
 To update to a specific upstream git tag or commit, use
 
   ./mach vendor aom -r <commit>
 
 The upstream aom git repository is:
 
     https://aomedia.googlesource.com/aom
 
-The git commit ID used was aadbb0251996c8ebb8310567bea330ab7ae9abe4 (Fri Jun 02 22:24:14 2017 +0000).
+The git commit ID used was f5bdeac22930ff4c6b219be49c843db35970b918 (Thu Jul 27 18:23:37 2017 +0000).
--- a/media/libaom/config/generic/aom_config.asm
+++ b/media/libaom/config/generic/aom_config.asm
@@ -63,16 +63,17 @@
 .equ CONFIG_INSPECTION ,  0
 .equ CONFIG_DECODE_PERF_TESTS ,  0
 .equ CONFIG_ENCODE_PERF_TESTS ,  0
 .equ CONFIG_COEFFICIENT_RANGE_CHECKING ,  0
 .equ CONFIG_LOWBITDEPTH ,  1
 .equ CONFIG_HIGHBITDEPTH ,  1
 .equ CONFIG_EXPERIMENTAL ,  0
 .equ CONFIG_SIZE_LIMIT ,  1
+.equ CONFIG_COLORSPACE_HEADERS ,  0
 .equ CONFIG_FP_MB_STATS ,  0
 .equ CONFIG_CDEF ,  1
 .equ CONFIG_VAR_TX ,  1
 .equ CONFIG_RECT_TX ,  1
 .equ CONFIG_RECT_TX_EXT ,  0
 .equ CONFIG_TPL_MV ,  0
 .equ CONFIG_DUAL_FILTER ,  1
 .equ CONFIG_CONVOLVE_ROUND ,  0
@@ -85,66 +86,82 @@
 .equ CONFIG_FILTER_INTRA ,  0
 .equ CONFIG_INTRA_EDGE ,  0
 .equ CONFIG_INTRABC ,  0
 .equ CONFIG_EXT_INTER ,  1
 .equ CONFIG_INTERINTRA ,  1
 .equ CONFIG_WEDGE ,  1
 .equ CONFIG_COMPOUND_SEGMENT ,  1
 .equ CONFIG_EXT_REFS ,  1
+.equ CONFIG_ALTREF2 ,  0
+.equ CONFIG_SPEED_REFS ,  0
+.equ CONFIG_GF_GROUPS ,  0
+.equ CONFIG_FLEX_REFS ,  0
 .equ CONFIG_GLOBAL_MOTION ,  1
 .equ CONFIG_NEW_QUANT ,  0
 .equ CONFIG_SUPERTX ,  0
 .equ CONFIG_ANS ,  0
 .equ CONFIG_LOOP_RESTORATION ,  0
 .equ CONFIG_EXT_PARTITION ,  0
 .equ CONFIG_EXT_PARTITION_TYPES ,  0
 .equ CONFIG_UNPOISON_PARTITION_CTX ,  0
 .equ CONFIG_EXT_TILE ,  0
 .equ CONFIG_MOTION_VAR ,  1
 .equ CONFIG_NCOBMC ,  0
 .equ CONFIG_WARPED_MOTION ,  1
 .equ CONFIG_Q_ADAPT_PROBS ,  0
 .equ CONFIG_BITSTREAM_DEBUG ,  0
+.equ CONFIG_INTER_STATS_ONLY ,  0
 .equ CONFIG_ALT_INTRA ,  1
 .equ CONFIG_PALETTE ,  1
 .equ CONFIG_PALETTE_DELTA_ENCODING ,  0
 .equ CONFIG_RAWBITS ,  0
 .equ CONFIG_EC_SMALLMUL ,  1
 .equ CONFIG_PVQ ,  0
 .equ CONFIG_CFL ,  0
 .equ CONFIG_XIPHRC ,  0
+.equ CONFIG_DCT_ONLY ,  0
+.equ CONFIG_DAALA_DCT4 ,  0
+.equ CONFIG_DAALA_DCT8 ,  0
 .equ CONFIG_CB4X4 ,  1
 .equ CONFIG_CHROMA_2X2 ,  0
 .equ CONFIG_CHROMA_SUB8X8 ,  1
 .equ CONFIG_FRAME_SIZE ,  0
 .equ CONFIG_DELTA_Q ,  1
 .equ CONFIG_EXT_DELTA_Q ,  1
 .equ CONFIG_ADAPT_SCAN ,  0
 .equ CONFIG_FILTER_7BIT ,  1
-.equ CONFIG_PARALLEL_DEBLOCKING ,  0
-.equ CONFIG_PARALLEL_DEBLOCKING_15TAP ,  0
+.equ CONFIG_PARALLEL_DEBLOCKING ,  1
 .equ CONFIG_LOOPFILTERING_ACROSS_TILES ,  1
-.equ CONFIG_TILE_GROUPS ,  1
-.equ CONFIG_EC_ADAPT ,  1
 .equ CONFIG_TEMPMV_SIGNALING ,  1
 .equ CONFIG_RD_DEBUG ,  0
 .equ CONFIG_REFERENCE_BUFFER ,  1
 .equ CONFIG_COEF_INTERLEAVE ,  0
 .equ CONFIG_ENTROPY_STATS ,  0
 .equ CONFIG_MASKED_TX ,  0
 .equ CONFIG_DEPENDENT_HORZTILES ,  0
+.equ CONFIG_DIST_8X8 ,  0
 .equ CONFIG_DAALA_DIST ,  0
 .equ CONFIG_TRIPRED ,  0
 .equ CONFIG_PALETTE_THROUGHPUT ,  1
 .equ CONFIG_REF_ADAPT ,  0
 .equ CONFIG_LV_MAP ,  0
 .equ CONFIG_TXK_SEL ,  0
 .equ CONFIG_MV_COMPRESS ,  1
 .equ CONFIG_FRAME_SUPERRES ,  0
 .equ CONFIG_NEW_MULTISYMBOL ,  0
 .equ CONFIG_COMPOUND_SINGLEREF ,  0
 .equ CONFIG_AOM_QM ,  0
 .equ CONFIG_ONE_SIDED_COMPOUND ,  1
-.equ CONFIG_SMOOTH_HV ,  0
+.equ CONFIG_EXT_COMP_REFS ,  0
+.equ CONFIG_SMOOTH_HV ,  1
 .equ CONFIG_VAR_REFS ,  0
+.equ CONFIG_RECT_INTRA_PRED ,  1
+.equ CONFIG_LGT ,  0
+.equ CONFIG_SBL_SYMBOL ,  0
+.equ CONFIG_NCOBMC_ADAPT_WEIGHT ,  0
+.equ CONFIG_BGSPRITE ,  0
+.equ CONFIG_VAR_TX_NO_TX_MODE ,  0
+.equ CONFIG_MRC_TX ,  0
+.equ CONFIG_LPF_DIRECT ,  0
+.equ CONFIG_UV_LVL ,  0
 .equ CONFIG_ANALYZER ,  0
 	.section	.note.GNU-stack,"",%progbits
--- a/media/libaom/config/generic/aom_config.h
+++ b/media/libaom/config/generic/aom_config.h
@@ -4,17 +4,17 @@
 /* the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License */
 /* was not distributed with this source code in the LICENSE file, you can */
 /* obtain it at www.aomedia.org/license/software. If the Alliance for Open */
 /* Media Patent License 1.0 was not distributed with this source code in the */
 /* PATENTS file, you can obtain it at www.aomedia.org/license/patent. */
 /* This file automatically generated by configure. Do not edit! */
 #ifndef AOM_CONFIG_H
 #define AOM_CONFIG_H
-#define RESTRICT
+#define RESTRICT    
 #define INLINE      inline
 #define ARCH_ARM 0
 #define ARCH_MIPS 0
 #define ARCH_X86 0
 #define ARCH_X86_64 0
 #define HAVE_NEON 0
 #define HAVE_NEON_ASM 0
 #define HAVE_MIPS32 0
@@ -73,16 +73,17 @@
 #define CONFIG_INSPECTION 0
 #define CONFIG_DECODE_PERF_TESTS 0
 #define CONFIG_ENCODE_PERF_TESTS 0
 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0
 #define CONFIG_LOWBITDEPTH 1
 #define CONFIG_HIGHBITDEPTH 1
 #define CONFIG_EXPERIMENTAL 0
 #define CONFIG_SIZE_LIMIT 1
+#define CONFIG_COLORSPACE_HEADERS 0
 #define CONFIG_FP_MB_STATS 0
 #define CONFIG_CDEF 1
 #define CONFIG_VAR_TX 1
 #define CONFIG_RECT_TX 1
 #define CONFIG_RECT_TX_EXT 0
 #define CONFIG_TPL_MV 0
 #define CONFIG_DUAL_FILTER 1
 #define CONFIG_CONVOLVE_ROUND 0
@@ -95,68 +96,84 @@
 #define CONFIG_FILTER_INTRA 0
 #define CONFIG_INTRA_EDGE 0
 #define CONFIG_INTRABC 0
 #define CONFIG_EXT_INTER 1
 #define CONFIG_INTERINTRA 1
 #define CONFIG_WEDGE 1
 #define CONFIG_COMPOUND_SEGMENT 1
 #define CONFIG_EXT_REFS 1
+#define CONFIG_ALTREF2 0
+#define CONFIG_SPEED_REFS 0
+#define CONFIG_GF_GROUPS 0
+#define CONFIG_FLEX_REFS 0
 #define CONFIG_GLOBAL_MOTION 1
 #define CONFIG_NEW_QUANT 0
 #define CONFIG_SUPERTX 0
 #define CONFIG_ANS 0
 #define CONFIG_LOOP_RESTORATION 0
 #define CONFIG_EXT_PARTITION 0
 #define CONFIG_EXT_PARTITION_TYPES 0
 #define CONFIG_UNPOISON_PARTITION_CTX 0
 #define CONFIG_EXT_TILE 0
 #define CONFIG_MOTION_VAR 1
 #define CONFIG_NCOBMC 0
 #define CONFIG_WARPED_MOTION 1
 #define CONFIG_Q_ADAPT_PROBS 0
 #define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_INTER_STATS_ONLY 0
 #define CONFIG_ALT_INTRA 1
 #define CONFIG_PALETTE 1
 #define CONFIG_PALETTE_DELTA_ENCODING 0
 #define CONFIG_RAWBITS 0
 #define CONFIG_EC_SMALLMUL 1
 #define CONFIG_PVQ 0
 #define CONFIG_CFL 0
 #define CONFIG_XIPHRC 0
+#define CONFIG_DCT_ONLY 0
+#define CONFIG_DAALA_DCT4 0
+#define CONFIG_DAALA_DCT8 0
 #define CONFIG_CB4X4 1
 #define CONFIG_CHROMA_2X2 0
 #define CONFIG_CHROMA_SUB8X8 1
 #define CONFIG_FRAME_SIZE 0
 #define CONFIG_DELTA_Q 1
 #define CONFIG_EXT_DELTA_Q 1
 #define CONFIG_ADAPT_SCAN 0
 #define CONFIG_FILTER_7BIT 1
-#define CONFIG_PARALLEL_DEBLOCKING 0
-#define CONFIG_PARALLEL_DEBLOCKING_15TAP 0
+#define CONFIG_PARALLEL_DEBLOCKING 1
 #define CONFIG_LOOPFILTERING_ACROSS_TILES 1
-#define CONFIG_TILE_GROUPS 1
-#define CONFIG_EC_ADAPT 1
 #define CONFIG_TEMPMV_SIGNALING 1
 #define CONFIG_RD_DEBUG 0
 #define CONFIG_REFERENCE_BUFFER 1
 #define CONFIG_COEF_INTERLEAVE 0
 #define CONFIG_ENTROPY_STATS 0
 #define CONFIG_MASKED_TX 0
 #define CONFIG_DEPENDENT_HORZTILES 0
+#define CONFIG_DIST_8X8 0
 #define CONFIG_DAALA_DIST 0
 #define CONFIG_TRIPRED 0
 #define CONFIG_PALETTE_THROUGHPUT 1
 #define CONFIG_REF_ADAPT 0
 #define CONFIG_LV_MAP 0
 #define CONFIG_TXK_SEL 0
 #define CONFIG_MV_COMPRESS 1
 #define CONFIG_FRAME_SUPERRES 0
 #define CONFIG_NEW_MULTISYMBOL 0
 #define CONFIG_COMPOUND_SINGLEREF 0
 #define CONFIG_AOM_QM 0
 #define CONFIG_ONE_SIDED_COMPOUND 1
-#define CONFIG_SMOOTH_HV 0
+#define CONFIG_EXT_COMP_REFS 0
+#define CONFIG_SMOOTH_HV 1
 #define CONFIG_VAR_REFS 0
+#define CONFIG_RECT_INTRA_PRED 1
+#define CONFIG_LGT 0
+#define CONFIG_SBL_SYMBOL 0
+#define CONFIG_NCOBMC_ADAPT_WEIGHT 0
+#define CONFIG_BGSPRITE 0
+#define CONFIG_VAR_TX_NO_TX_MODE 0
+#define CONFIG_MRC_TX 0
+#define CONFIG_LPF_DIRECT 0
+#define CONFIG_UV_LVL 0
 #define CONFIG_ANALYZER 0
 #define DECODE_WIDTH_LIMIT 8192
 #define DECODE_HEIGHT_LIMIT 4608
 #endif /* AOM_CONFIG_H */
--- a/media/libaom/config/generic/aom_dsp_rtcd.h
+++ b/media/libaom/config/generic/aom_dsp_rtcd.h
@@ -27,250 +27,457 @@ void aom_blend_a64_mask_c(uint8_t *dst, 
 #define aom_blend_a64_mask aom_blend_a64_mask_c
 
 void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w);
 #define aom_blend_a64_vmask aom_blend_a64_vmask_c
 
 void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
 #define aom_comp_avg_pred aom_comp_avg_pred_c
 
-void aom_comp_avg_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+void aom_comp_avg_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref, int ref_stride);
 #define aom_comp_avg_upsampled_pred aom_comp_avg_upsampled_pred_c
 
 void aom_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 #define aom_comp_mask_pred aom_comp_mask_pred_c
 
-void aom_comp_mask_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_comp_mask_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 #define aom_comp_mask_upsampled_pred aom_comp_mask_upsampled_pred_c
 
 void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 #define aom_convolve8 aom_convolve8_c
 
 void aom_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 #define aom_convolve8_avg aom_convolve8_avg_c
 
 void aom_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 #define aom_convolve8_avg_horiz aom_convolve8_avg_horiz_c
 
+void aom_convolve8_avg_horiz_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_avg_horiz_scale aom_convolve8_avg_horiz_scale_c
+
+void aom_convolve8_avg_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_avg_scale aom_convolve8_avg_scale_c
+
 void aom_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 #define aom_convolve8_avg_vert aom_convolve8_avg_vert_c
 
+void aom_convolve8_avg_vert_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_avg_vert_scale aom_convolve8_avg_vert_scale_c
+
 void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 #define aom_convolve8_horiz aom_convolve8_horiz_c
 
+void aom_convolve8_horiz_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_horiz_scale aom_convolve8_horiz_scale_c
+
+void aom_convolve8_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_scale aom_convolve8_scale_c
+
 void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 #define aom_convolve8_vert aom_convolve8_vert_c
 
+void aom_convolve8_vert_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_vert_scale aom_convolve8_vert_scale_c
+
 void aom_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 #define aom_convolve_avg aom_convolve_avg_c
 
 void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 #define aom_convolve_copy aom_convolve_copy_c
 
 void aom_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d117_predictor_16x16 aom_d117_predictor_16x16_c
 
+void aom_d117_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_16x32 aom_d117_predictor_16x32_c
+
+void aom_d117_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_16x8 aom_d117_predictor_16x8_c
+
 void aom_d117_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d117_predictor_2x2 aom_d117_predictor_2x2_c
 
+void aom_d117_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_32x16 aom_d117_predictor_32x16_c
+
 void aom_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d117_predictor_32x32 aom_d117_predictor_32x32_c
 
 void aom_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d117_predictor_4x4 aom_d117_predictor_4x4_c
 
+void aom_d117_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_4x8 aom_d117_predictor_4x8_c
+
+void aom_d117_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_8x16 aom_d117_predictor_8x16_c
+
+void aom_d117_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_8x4 aom_d117_predictor_8x4_c
+
 void aom_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d117_predictor_8x8 aom_d117_predictor_8x8_c
 
 void aom_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d135_predictor_16x16 aom_d135_predictor_16x16_c
 
+void aom_d135_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_16x32 aom_d135_predictor_16x32_c
+
+void aom_d135_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_16x8 aom_d135_predictor_16x8_c
+
 void aom_d135_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d135_predictor_2x2 aom_d135_predictor_2x2_c
 
+void aom_d135_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_32x16 aom_d135_predictor_32x16_c
+
 void aom_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d135_predictor_32x32 aom_d135_predictor_32x32_c
 
 void aom_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d135_predictor_4x4 aom_d135_predictor_4x4_c
 
+void aom_d135_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_4x8 aom_d135_predictor_4x8_c
+
+void aom_d135_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_8x16 aom_d135_predictor_8x16_c
+
+void aom_d135_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_8x4 aom_d135_predictor_8x4_c
+
 void aom_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d135_predictor_8x8 aom_d135_predictor_8x8_c
 
 void aom_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d153_predictor_16x16 aom_d153_predictor_16x16_c
 
+void aom_d153_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_16x32 aom_d153_predictor_16x32_c
+
+void aom_d153_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_16x8 aom_d153_predictor_16x8_c
+
 void aom_d153_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d153_predictor_2x2 aom_d153_predictor_2x2_c
 
+void aom_d153_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_32x16 aom_d153_predictor_32x16_c
+
 void aom_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d153_predictor_32x32 aom_d153_predictor_32x32_c
 
 void aom_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d153_predictor_4x4 aom_d153_predictor_4x4_c
 
+void aom_d153_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_4x8 aom_d153_predictor_4x8_c
+
+void aom_d153_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_8x16 aom_d153_predictor_8x16_c
+
+void aom_d153_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_8x4 aom_d153_predictor_8x4_c
+
 void aom_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d153_predictor_8x8 aom_d153_predictor_8x8_c
 
 void aom_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d207e_predictor_16x16 aom_d207e_predictor_16x16_c
 
+void aom_d207e_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_16x32 aom_d207e_predictor_16x32_c
+
+void aom_d207e_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_16x8 aom_d207e_predictor_16x8_c
+
 void aom_d207e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d207e_predictor_2x2 aom_d207e_predictor_2x2_c
 
+void aom_d207e_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_32x16 aom_d207e_predictor_32x16_c
+
 void aom_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d207e_predictor_32x32 aom_d207e_predictor_32x32_c
 
 void aom_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d207e_predictor_4x4 aom_d207e_predictor_4x4_c
 
+void aom_d207e_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_4x8 aom_d207e_predictor_4x8_c
+
+void aom_d207e_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_8x16 aom_d207e_predictor_8x16_c
+
+void aom_d207e_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_8x4 aom_d207e_predictor_8x4_c
+
 void aom_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d207e_predictor_8x8 aom_d207e_predictor_8x8_c
 
 void aom_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d45e_predictor_16x16 aom_d45e_predictor_16x16_c
 
+void aom_d45e_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_16x32 aom_d45e_predictor_16x32_c
+
+void aom_d45e_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_16x8 aom_d45e_predictor_16x8_c
+
 void aom_d45e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d45e_predictor_2x2 aom_d45e_predictor_2x2_c
 
+void aom_d45e_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_32x16 aom_d45e_predictor_32x16_c
+
 void aom_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d45e_predictor_32x32 aom_d45e_predictor_32x32_c
 
 void aom_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d45e_predictor_4x4 aom_d45e_predictor_4x4_c
 
+void aom_d45e_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_4x8 aom_d45e_predictor_4x8_c
+
+void aom_d45e_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_8x16 aom_d45e_predictor_8x16_c
+
+void aom_d45e_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_8x4 aom_d45e_predictor_8x4_c
+
 void aom_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d45e_predictor_8x8 aom_d45e_predictor_8x8_c
 
 void aom_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d63e_predictor_16x16 aom_d63e_predictor_16x16_c
 
+void aom_d63e_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_16x32 aom_d63e_predictor_16x32_c
+
+void aom_d63e_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_16x8 aom_d63e_predictor_16x8_c
+
 void aom_d63e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d63e_predictor_2x2 aom_d63e_predictor_2x2_c
 
+void aom_d63e_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_32x16 aom_d63e_predictor_32x16_c
+
 void aom_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d63e_predictor_32x32 aom_d63e_predictor_32x32_c
 
 void aom_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d63e_predictor_4x4 aom_d63e_predictor_4x4_c
 
+void aom_d63e_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_4x8 aom_d63e_predictor_4x8_c
+
+void aom_d63e_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_8x16 aom_d63e_predictor_8x16_c
+
+void aom_d63e_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_8x4 aom_d63e_predictor_8x4_c
+
 void aom_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d63e_predictor_8x8 aom_d63e_predictor_8x8_c
 
 void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_128_predictor_16x16 aom_dc_128_predictor_16x16_c
 
+void aom_dc_128_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_c
+
+void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c
+
 void aom_dc_128_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
 
+void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_32x16 aom_dc_128_predictor_32x16_c
+
 void aom_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_128_predictor_32x32 aom_dc_128_predictor_32x32_c
 
 void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_128_predictor_4x4 aom_dc_128_predictor_4x4_c
 
+void aom_dc_128_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_c
+
+void aom_dc_128_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_c
+
+void aom_dc_128_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x4 aom_dc_128_predictor_8x4_c
+
 void aom_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_128_predictor_8x8 aom_dc_128_predictor_8x8_c
 
 void aom_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_left_predictor_16x16 aom_dc_left_predictor_16x16_c
 
+void aom_dc_left_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_c
+
+void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c
+
 void aom_dc_left_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
 
+void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_32x16 aom_dc_left_predictor_32x16_c
+
 void aom_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_left_predictor_32x32 aom_dc_left_predictor_32x32_c
 
 void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_left_predictor_4x4 aom_dc_left_predictor_4x4_c
 
+void aom_dc_left_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_c
+
+void aom_dc_left_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_c
+
+void aom_dc_left_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x4 aom_dc_left_predictor_8x4_c
+
 void aom_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_left_predictor_8x8 aom_dc_left_predictor_8x8_c
 
 void aom_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_predictor_16x16 aom_dc_predictor_16x16_c
 
+void aom_dc_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x32 aom_dc_predictor_16x32_c
+
+void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c
+
 void aom_dc_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
 
+void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_32x16 aom_dc_predictor_32x16_c
+
 void aom_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_predictor_32x32 aom_dc_predictor_32x32_c
 
 void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_predictor_4x4 aom_dc_predictor_4x4_c
 
+void aom_dc_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x8 aom_dc_predictor_4x8_c
+
+void aom_dc_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x16 aom_dc_predictor_8x16_c
+
+void aom_dc_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x4 aom_dc_predictor_8x4_c
+
 void aom_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_predictor_8x8 aom_dc_predictor_8x8_c
 
 void aom_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_top_predictor_16x16 aom_dc_top_predictor_16x16_c
 
+void aom_dc_top_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_c
+
+void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c
+
 void aom_dc_top_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
 
+void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_32x16 aom_dc_top_predictor_32x16_c
+
 void aom_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_top_predictor_32x32 aom_dc_top_predictor_32x32_c
 
 void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_top_predictor_4x4 aom_dc_top_predictor_4x4_c
 
+void aom_dc_top_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_c
+
+void aom_dc_top_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_c
+
+void aom_dc_top_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x4 aom_dc_top_predictor_8x4_c
+
 void aom_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_top_predictor_8x8 aom_dc_top_predictor_8x8_c
 
 void aom_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_fdct16x16 aom_fdct16x16_c
 
-void aom_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
-#define aom_fdct16x16_1 aom_fdct16x16_1_c
-
 void aom_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_fdct32x32 aom_fdct32x32_c
 
-void aom_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
-#define aom_fdct32x32_1 aom_fdct32x32_1_c
-
 void aom_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_fdct32x32_rd aom_fdct32x32_rd_c
 
 void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_fdct4x4 aom_fdct4x4_c
 
 void aom_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_fdct4x4_1 aom_fdct4x4_1_c
 
 void aom_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_fdct8x8 aom_fdct8x8_c
 
-void aom_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
-#define aom_fdct8x8_1 aom_fdct8x8_1_c
-
 void aom_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
 #define aom_get16x16var aom_get16x16var_c
 
 unsigned int aom_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride);
 #define aom_get4x4sse_cs aom_get4x4sse_cs_c
 
 void aom_get8x8var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
 #define aom_get8x8var aom_get8x8var_c
 
 unsigned int aom_get_mb_ss_c(const int16_t *);
 #define aom_get_mb_ss aom_get_mb_ss_c
 
 void aom_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_h_predictor_16x16 aom_h_predictor_16x16_c
 
+void aom_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x32 aom_h_predictor_16x32_c
+
+void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x8 aom_h_predictor_16x8_c
+
 void aom_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_h_predictor_2x2 aom_h_predictor_2x2_c
 
+void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x16 aom_h_predictor_32x16_c
+
 void aom_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_h_predictor_32x32 aom_h_predictor_32x32_c
 
 void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_h_predictor_4x4 aom_h_predictor_4x4_c
 
+void aom_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x8 aom_h_predictor_4x8_c
+
+void aom_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x16 aom_h_predictor_8x16_c
+
+void aom_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x4 aom_h_predictor_8x4_c
+
 void aom_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_h_predictor_8x8 aom_h_predictor_8x8_c
 
 void aom_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
 #define aom_hadamard_16x16 aom_hadamard_16x16_c
 
 void aom_hadamard_8x8_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
 #define aom_hadamard_8x8 aom_hadamard_8x8_c
@@ -282,43 +489,55 @@ void aom_highbd_10_get8x8var_c(const uin
 #define aom_highbd_10_get8x8var aom_highbd_10_get8x8var_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance16x16 aom_highbd_10_masked_sub_pixel_variance16x16_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance16x32 aom_highbd_10_masked_sub_pixel_variance16x32_c
 
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance16x4 aom_highbd_10_masked_sub_pixel_variance16x4_c
+
 unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance16x8 aom_highbd_10_masked_sub_pixel_variance16x8_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance32x16 aom_highbd_10_masked_sub_pixel_variance32x16_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance32x32 aom_highbd_10_masked_sub_pixel_variance32x32_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance32x64 aom_highbd_10_masked_sub_pixel_variance32x64_c
 
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance32x8 aom_highbd_10_masked_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance4x16 aom_highbd_10_masked_sub_pixel_variance4x16_c
+
 unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance4x4 aom_highbd_10_masked_sub_pixel_variance4x4_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance4x8 aom_highbd_10_masked_sub_pixel_variance4x8_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance64x32 aom_highbd_10_masked_sub_pixel_variance64x32_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance64x64 aom_highbd_10_masked_sub_pixel_variance64x64_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance8x16 aom_highbd_10_masked_sub_pixel_variance8x16_c
 
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance8x32 aom_highbd_10_masked_sub_pixel_variance8x32_c
+
 unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance8x4 aom_highbd_10_masked_sub_pixel_variance8x4_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance8x8 aom_highbd_10_masked_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define aom_highbd_10_mse16x16 aom_highbd_10_mse16x16_c
@@ -333,172 +552,223 @@ unsigned int aom_highbd_10_mse8x8_c(cons
 #define aom_highbd_10_mse8x8 aom_highbd_10_mse8x8_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance16x16 aom_highbd_10_obmc_sub_pixel_variance16x16_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance16x32 aom_highbd_10_obmc_sub_pixel_variance16x32_c
 
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x4 aom_highbd_10_obmc_sub_pixel_variance16x4_c
+
 unsigned int aom_highbd_10_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance16x8 aom_highbd_10_obmc_sub_pixel_variance16x8_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance32x16 aom_highbd_10_obmc_sub_pixel_variance32x16_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance32x32 aom_highbd_10_obmc_sub_pixel_variance32x32_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance32x64 aom_highbd_10_obmc_sub_pixel_variance32x64_c
 
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x8 aom_highbd_10_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x16 aom_highbd_10_obmc_sub_pixel_variance4x16_c
+
 unsigned int aom_highbd_10_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance4x4 aom_highbd_10_obmc_sub_pixel_variance4x4_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance4x8 aom_highbd_10_obmc_sub_pixel_variance4x8_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance64x32 aom_highbd_10_obmc_sub_pixel_variance64x32_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance64x64 aom_highbd_10_obmc_sub_pixel_variance64x64_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance8x16 aom_highbd_10_obmc_sub_pixel_variance8x16_c
 
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x32 aom_highbd_10_obmc_sub_pixel_variance8x32_c
+
 unsigned int aom_highbd_10_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance8x4 aom_highbd_10_obmc_sub_pixel_variance8x4_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance8x8 aom_highbd_10_obmc_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_10_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance16x16 aom_highbd_10_obmc_variance16x16_c
 
 unsigned int aom_highbd_10_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance16x32 aom_highbd_10_obmc_variance16x32_c
 
+unsigned int aom_highbd_10_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance16x4 aom_highbd_10_obmc_variance16x4_c
+
 unsigned int aom_highbd_10_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance16x8 aom_highbd_10_obmc_variance16x8_c
 
 unsigned int aom_highbd_10_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance32x16 aom_highbd_10_obmc_variance32x16_c
 
 unsigned int aom_highbd_10_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance32x32 aom_highbd_10_obmc_variance32x32_c
 
 unsigned int aom_highbd_10_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance32x64 aom_highbd_10_obmc_variance32x64_c
 
+unsigned int aom_highbd_10_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance32x8 aom_highbd_10_obmc_variance32x8_c
+
+unsigned int aom_highbd_10_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance4x16 aom_highbd_10_obmc_variance4x16_c
+
 unsigned int aom_highbd_10_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance4x4 aom_highbd_10_obmc_variance4x4_c
 
 unsigned int aom_highbd_10_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance4x8 aom_highbd_10_obmc_variance4x8_c
 
 unsigned int aom_highbd_10_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance64x32 aom_highbd_10_obmc_variance64x32_c
 
 unsigned int aom_highbd_10_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance64x64 aom_highbd_10_obmc_variance64x64_c
 
 unsigned int aom_highbd_10_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance8x16 aom_highbd_10_obmc_variance8x16_c
 
+unsigned int aom_highbd_10_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance8x32 aom_highbd_10_obmc_variance8x32_c
+
 unsigned int aom_highbd_10_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance8x4 aom_highbd_10_obmc_variance8x4_c
 
 unsigned int aom_highbd_10_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance8x8 aom_highbd_10_obmc_variance8x8_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance16x16 aom_highbd_10_sub_pixel_avg_variance16x16_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance16x32 aom_highbd_10_sub_pixel_avg_variance16x32_c
 
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x4 aom_highbd_10_sub_pixel_avg_variance16x4_c
+
 uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance16x8 aom_highbd_10_sub_pixel_avg_variance16x8_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance32x16 aom_highbd_10_sub_pixel_avg_variance32x16_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance32x32 aom_highbd_10_sub_pixel_avg_variance32x32_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance32x64 aom_highbd_10_sub_pixel_avg_variance32x64_c
 
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x8 aom_highbd_10_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x16 aom_highbd_10_sub_pixel_avg_variance4x16_c
+
 uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance4x4 aom_highbd_10_sub_pixel_avg_variance4x4_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance4x8 aom_highbd_10_sub_pixel_avg_variance4x8_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance64x32 aom_highbd_10_sub_pixel_avg_variance64x32_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance64x64 aom_highbd_10_sub_pixel_avg_variance64x64_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance8x16 aom_highbd_10_sub_pixel_avg_variance8x16_c
 
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x32 aom_highbd_10_sub_pixel_avg_variance8x32_c
+
 uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance8x4 aom_highbd_10_sub_pixel_avg_variance8x4_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance8x8 aom_highbd_10_sub_pixel_avg_variance8x8_c
 
 uint32_t aom_highbd_10_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance16x16 aom_highbd_10_sub_pixel_variance16x16_c
 
 uint32_t aom_highbd_10_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance16x32 aom_highbd_10_sub_pixel_variance16x32_c
 
+uint32_t aom_highbd_10_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x4 aom_highbd_10_sub_pixel_variance16x4_c
+
 uint32_t aom_highbd_10_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance16x8 aom_highbd_10_sub_pixel_variance16x8_c
 
 uint32_t aom_highbd_10_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance32x16 aom_highbd_10_sub_pixel_variance32x16_c
 
 uint32_t aom_highbd_10_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance32x32 aom_highbd_10_sub_pixel_variance32x32_c
 
 uint32_t aom_highbd_10_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance32x64 aom_highbd_10_sub_pixel_variance32x64_c
 
+uint32_t aom_highbd_10_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance32x8 aom_highbd_10_sub_pixel_variance32x8_c
+
+uint32_t aom_highbd_10_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x16 aom_highbd_10_sub_pixel_variance4x16_c
+
 uint32_t aom_highbd_10_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance4x4 aom_highbd_10_sub_pixel_variance4x4_c
 
 uint32_t aom_highbd_10_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance4x8 aom_highbd_10_sub_pixel_variance4x8_c
 
 uint32_t aom_highbd_10_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance64x32 aom_highbd_10_sub_pixel_variance64x32_c
 
 uint32_t aom_highbd_10_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance64x64 aom_highbd_10_sub_pixel_variance64x64_c
 
 uint32_t aom_highbd_10_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance8x16 aom_highbd_10_sub_pixel_variance8x16_c
 
+uint32_t aom_highbd_10_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x32 aom_highbd_10_sub_pixel_variance8x32_c
+
 uint32_t aom_highbd_10_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance8x4 aom_highbd_10_sub_pixel_variance8x4_c
 
 uint32_t aom_highbd_10_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance8x8 aom_highbd_10_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_10_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance16x16 aom_highbd_10_variance16x16_c
 
 unsigned int aom_highbd_10_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance16x32 aom_highbd_10_variance16x32_c
 
+unsigned int aom_highbd_10_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance16x4 aom_highbd_10_variance16x4_c
+
 unsigned int aom_highbd_10_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance16x8 aom_highbd_10_variance16x8_c
 
 unsigned int aom_highbd_10_variance2x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance2x2 aom_highbd_10_variance2x2_c
 
 unsigned int aom_highbd_10_variance2x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance2x4 aom_highbd_10_variance2x4_c
@@ -507,16 +777,22 @@ unsigned int aom_highbd_10_variance32x16
 #define aom_highbd_10_variance32x16 aom_highbd_10_variance32x16_c
 
 unsigned int aom_highbd_10_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance32x32 aom_highbd_10_variance32x32_c
 
 unsigned int aom_highbd_10_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance32x64 aom_highbd_10_variance32x64_c
 
+unsigned int aom_highbd_10_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance32x8 aom_highbd_10_variance32x8_c
+
+unsigned int aom_highbd_10_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x16 aom_highbd_10_variance4x16_c
+
 unsigned int aom_highbd_10_variance4x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance4x2 aom_highbd_10_variance4x2_c
 
 unsigned int aom_highbd_10_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance4x4 aom_highbd_10_variance4x4_c
 
 unsigned int aom_highbd_10_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance4x8 aom_highbd_10_variance4x8_c
@@ -525,16 +801,19 @@ unsigned int aom_highbd_10_variance64x32
 #define aom_highbd_10_variance64x32 aom_highbd_10_variance64x32_c
 
 unsigned int aom_highbd_10_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance64x64 aom_highbd_10_variance64x64_c
 
 unsigned int aom_highbd_10_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance8x16 aom_highbd_10_variance8x16_c
 
+unsigned int aom_highbd_10_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance8x32 aom_highbd_10_variance8x32_c
+
 unsigned int aom_highbd_10_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance8x4 aom_highbd_10_variance8x4_c
 
 unsigned int aom_highbd_10_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance8x8 aom_highbd_10_variance8x8_c
 
 void aom_highbd_12_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
 #define aom_highbd_12_get16x16var aom_highbd_12_get16x16var_c
@@ -543,43 +822,55 @@ void aom_highbd_12_get8x8var_c(const uin
 #define aom_highbd_12_get8x8var aom_highbd_12_get8x8var_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance16x16 aom_highbd_12_masked_sub_pixel_variance16x16_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance16x32 aom_highbd_12_masked_sub_pixel_variance16x32_c
 
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance16x4 aom_highbd_12_masked_sub_pixel_variance16x4_c
+
 unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance16x8 aom_highbd_12_masked_sub_pixel_variance16x8_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance32x16 aom_highbd_12_masked_sub_pixel_variance32x16_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance32x32 aom_highbd_12_masked_sub_pixel_variance32x32_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance32x64 aom_highbd_12_masked_sub_pixel_variance32x64_c
 
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance32x8 aom_highbd_12_masked_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance4x16 aom_highbd_12_masked_sub_pixel_variance4x16_c
+
 unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance4x4 aom_highbd_12_masked_sub_pixel_variance4x4_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance4x8 aom_highbd_12_masked_sub_pixel_variance4x8_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance64x32 aom_highbd_12_masked_sub_pixel_variance64x32_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance64x64 aom_highbd_12_masked_sub_pixel_variance64x64_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance8x16 aom_highbd_12_masked_sub_pixel_variance8x16_c
 
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance8x32 aom_highbd_12_masked_sub_pixel_variance8x32_c
+
 unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance8x4 aom_highbd_12_masked_sub_pixel_variance8x4_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance8x8 aom_highbd_12_masked_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_12_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define aom_highbd_12_mse16x16 aom_highbd_12_mse16x16_c
@@ -594,172 +885,223 @@ unsigned int aom_highbd_12_mse8x8_c(cons
 #define aom_highbd_12_mse8x8 aom_highbd_12_mse8x8_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance16x16 aom_highbd_12_obmc_sub_pixel_variance16x16_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance16x32 aom_highbd_12_obmc_sub_pixel_variance16x32_c
 
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x4 aom_highbd_12_obmc_sub_pixel_variance16x4_c
+
 unsigned int aom_highbd_12_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance16x8 aom_highbd_12_obmc_sub_pixel_variance16x8_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance32x16 aom_highbd_12_obmc_sub_pixel_variance32x16_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance32x32 aom_highbd_12_obmc_sub_pixel_variance32x32_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance32x64 aom_highbd_12_obmc_sub_pixel_variance32x64_c
 
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x8 aom_highbd_12_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x16 aom_highbd_12_obmc_sub_pixel_variance4x16_c
+
 unsigned int aom_highbd_12_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance4x4 aom_highbd_12_obmc_sub_pixel_variance4x4_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance4x8 aom_highbd_12_obmc_sub_pixel_variance4x8_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance64x32 aom_highbd_12_obmc_sub_pixel_variance64x32_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance64x64 aom_highbd_12_obmc_sub_pixel_variance64x64_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance8x16 aom_highbd_12_obmc_sub_pixel_variance8x16_c
 
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x32 aom_highbd_12_obmc_sub_pixel_variance8x32_c
+
 unsigned int aom_highbd_12_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance8x4 aom_highbd_12_obmc_sub_pixel_variance8x4_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance8x8 aom_highbd_12_obmc_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_12_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance16x16 aom_highbd_12_obmc_variance16x16_c
 
 unsigned int aom_highbd_12_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance16x32 aom_highbd_12_obmc_variance16x32_c
 
+unsigned int aom_highbd_12_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance16x4 aom_highbd_12_obmc_variance16x4_c
+
 unsigned int aom_highbd_12_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance16x8 aom_highbd_12_obmc_variance16x8_c
 
 unsigned int aom_highbd_12_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance32x16 aom_highbd_12_obmc_variance32x16_c
 
 unsigned int aom_highbd_12_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance32x32 aom_highbd_12_obmc_variance32x32_c
 
 unsigned int aom_highbd_12_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance32x64 aom_highbd_12_obmc_variance32x64_c
 
+unsigned int aom_highbd_12_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance32x8 aom_highbd_12_obmc_variance32x8_c
+
+unsigned int aom_highbd_12_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance4x16 aom_highbd_12_obmc_variance4x16_c
+
 unsigned int aom_highbd_12_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance4x4 aom_highbd_12_obmc_variance4x4_c
 
 unsigned int aom_highbd_12_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance4x8 aom_highbd_12_obmc_variance4x8_c
 
 unsigned int aom_highbd_12_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance64x32 aom_highbd_12_obmc_variance64x32_c
 
 unsigned int aom_highbd_12_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance64x64 aom_highbd_12_obmc_variance64x64_c
 
 unsigned int aom_highbd_12_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance8x16 aom_highbd_12_obmc_variance8x16_c
 
+unsigned int aom_highbd_12_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance8x32 aom_highbd_12_obmc_variance8x32_c
+
 unsigned int aom_highbd_12_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance8x4 aom_highbd_12_obmc_variance8x4_c
 
 unsigned int aom_highbd_12_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance8x8 aom_highbd_12_obmc_variance8x8_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance16x16 aom_highbd_12_sub_pixel_avg_variance16x16_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance16x32 aom_highbd_12_sub_pixel_avg_variance16x32_c
 
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x4 aom_highbd_12_sub_pixel_avg_variance16x4_c
+
 uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance16x8 aom_highbd_12_sub_pixel_avg_variance16x8_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance32x16 aom_highbd_12_sub_pixel_avg_variance32x16_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance32x32 aom_highbd_12_sub_pixel_avg_variance32x32_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance32x64 aom_highbd_12_sub_pixel_avg_variance32x64_c
 
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x8 aom_highbd_12_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x16 aom_highbd_12_sub_pixel_avg_variance4x16_c
+
 uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance4x4 aom_highbd_12_sub_pixel_avg_variance4x4_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance4x8 aom_highbd_12_sub_pixel_avg_variance4x8_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance64x32 aom_highbd_12_sub_pixel_avg_variance64x32_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance64x64 aom_highbd_12_sub_pixel_avg_variance64x64_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance8x16 aom_highbd_12_sub_pixel_avg_variance8x16_c
 
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x32 aom_highbd_12_sub_pixel_avg_variance8x32_c
+
 uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance8x4 aom_highbd_12_sub_pixel_avg_variance8x4_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance8x8 aom_highbd_12_sub_pixel_avg_variance8x8_c
 
 uint32_t aom_highbd_12_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance16x16 aom_highbd_12_sub_pixel_variance16x16_c
 
 uint32_t aom_highbd_12_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance16x32 aom_highbd_12_sub_pixel_variance16x32_c
 
+uint32_t aom_highbd_12_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x4 aom_highbd_12_sub_pixel_variance16x4_c
+
 uint32_t aom_highbd_12_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance16x8 aom_highbd_12_sub_pixel_variance16x8_c
 
 uint32_t aom_highbd_12_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance32x16 aom_highbd_12_sub_pixel_variance32x16_c
 
 uint32_t aom_highbd_12_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance32x32 aom_highbd_12_sub_pixel_variance32x32_c
 
 uint32_t aom_highbd_12_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance32x64 aom_highbd_12_sub_pixel_variance32x64_c
 
+uint32_t aom_highbd_12_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x8 aom_highbd_12_sub_pixel_variance32x8_c
+
+uint32_t aom_highbd_12_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x16 aom_highbd_12_sub_pixel_variance4x16_c
+
 uint32_t aom_highbd_12_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance4x4 aom_highbd_12_sub_pixel_variance4x4_c
 
 uint32_t aom_highbd_12_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance4x8 aom_highbd_12_sub_pixel_variance4x8_c
 
 uint32_t aom_highbd_12_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance64x32 aom_highbd_12_sub_pixel_variance64x32_c
 
 uint32_t aom_highbd_12_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance64x64 aom_highbd_12_sub_pixel_variance64x64_c
 
 uint32_t aom_highbd_12_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance8x16 aom_highbd_12_sub_pixel_variance8x16_c
 
+uint32_t aom_highbd_12_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x32 aom_highbd_12_sub_pixel_variance8x32_c
+
 uint32_t aom_highbd_12_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance8x4 aom_highbd_12_sub_pixel_variance8x4_c
 
 uint32_t aom_highbd_12_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance8x8 aom_highbd_12_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_12_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance16x16 aom_highbd_12_variance16x16_c
 
 unsigned int aom_highbd_12_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance16x32 aom_highbd_12_variance16x32_c
 
+unsigned int aom_highbd_12_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x4 aom_highbd_12_variance16x4_c
+
 unsigned int aom_highbd_12_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance16x8 aom_highbd_12_variance16x8_c
 
 unsigned int aom_highbd_12_variance2x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance2x2 aom_highbd_12_variance2x2_c
 
 unsigned int aom_highbd_12_variance2x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance2x4 aom_highbd_12_variance2x4_c
@@ -768,16 +1110,22 @@ unsigned int aom_highbd_12_variance32x16
 #define aom_highbd_12_variance32x16 aom_highbd_12_variance32x16_c
 
 unsigned int aom_highbd_12_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance32x32 aom_highbd_12_variance32x32_c
 
 unsigned int aom_highbd_12_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance32x64 aom_highbd_12_variance32x64_c
 
+unsigned int aom_highbd_12_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x8 aom_highbd_12_variance32x8_c
+
+unsigned int aom_highbd_12_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x16 aom_highbd_12_variance4x16_c
+
 unsigned int aom_highbd_12_variance4x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance4x2 aom_highbd_12_variance4x2_c
 
 unsigned int aom_highbd_12_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance4x4 aom_highbd_12_variance4x4_c
 
 unsigned int aom_highbd_12_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance4x8 aom_highbd_12_variance4x8_c
@@ -786,16 +1134,19 @@ unsigned int aom_highbd_12_variance64x32
 #define aom_highbd_12_variance64x32 aom_highbd_12_variance64x32_c
 
 unsigned int aom_highbd_12_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance64x64 aom_highbd_12_variance64x64_c
 
 unsigned int aom_highbd_12_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance8x16 aom_highbd_12_variance8x16_c
 
+unsigned int aom_highbd_12_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x32 aom_highbd_12_variance8x32_c
+
 unsigned int aom_highbd_12_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance8x4 aom_highbd_12_variance8x4_c
 
 unsigned int aom_highbd_12_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance8x8 aom_highbd_12_variance8x8_c
 
 void aom_highbd_8_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
 #define aom_highbd_8_get16x16var aom_highbd_8_get16x16var_c
@@ -804,43 +1155,55 @@ void aom_highbd_8_get8x8var_c(const uint
 #define aom_highbd_8_get8x8var aom_highbd_8_get8x8var_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance16x16 aom_highbd_8_masked_sub_pixel_variance16x16_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance16x32 aom_highbd_8_masked_sub_pixel_variance16x32_c
 
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance16x4 aom_highbd_8_masked_sub_pixel_variance16x4_c
+
 unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance16x8 aom_highbd_8_masked_sub_pixel_variance16x8_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance32x16 aom_highbd_8_masked_sub_pixel_variance32x16_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance32x32 aom_highbd_8_masked_sub_pixel_variance32x32_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance32x64 aom_highbd_8_masked_sub_pixel_variance32x64_c
 
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance32x8 aom_highbd_8_masked_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance4x16 aom_highbd_8_masked_sub_pixel_variance4x16_c
+
 unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance4x4 aom_highbd_8_masked_sub_pixel_variance4x4_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance4x8 aom_highbd_8_masked_sub_pixel_variance4x8_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance64x32 aom_highbd_8_masked_sub_pixel_variance64x32_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance64x64 aom_highbd_8_masked_sub_pixel_variance64x64_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance8x16 aom_highbd_8_masked_sub_pixel_variance8x16_c
 
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance8x32 aom_highbd_8_masked_sub_pixel_variance8x32_c
+
 unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance8x4 aom_highbd_8_masked_sub_pixel_variance8x4_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance8x8 aom_highbd_8_masked_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_8_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define aom_highbd_8_mse16x16 aom_highbd_8_mse16x16_c
@@ -855,94 +1218,121 @@ unsigned int aom_highbd_8_mse8x8_c(const
 #define aom_highbd_8_mse8x8 aom_highbd_8_mse8x8_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance16x16 aom_highbd_8_sub_pixel_avg_variance16x16_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance16x32 aom_highbd_8_sub_pixel_avg_variance16x32_c
 
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x4 aom_highbd_8_sub_pixel_avg_variance16x4_c
+
 uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance16x8 aom_highbd_8_sub_pixel_avg_variance16x8_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance32x16 aom_highbd_8_sub_pixel_avg_variance32x16_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance32x32 aom_highbd_8_sub_pixel_avg_variance32x32_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance32x64 aom_highbd_8_sub_pixel_avg_variance32x64_c
 
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x8 aom_highbd_8_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x16 aom_highbd_8_sub_pixel_avg_variance4x16_c
+
 uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance4x4 aom_highbd_8_sub_pixel_avg_variance4x4_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance4x8 aom_highbd_8_sub_pixel_avg_variance4x8_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance64x32 aom_highbd_8_sub_pixel_avg_variance64x32_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance64x64 aom_highbd_8_sub_pixel_avg_variance64x64_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance8x16 aom_highbd_8_sub_pixel_avg_variance8x16_c
 
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x32 aom_highbd_8_sub_pixel_avg_variance8x32_c
+
 uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance8x4 aom_highbd_8_sub_pixel_avg_variance8x4_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance8x8 aom_highbd_8_sub_pixel_avg_variance8x8_c
 
 uint32_t aom_highbd_8_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance16x16 aom_highbd_8_sub_pixel_variance16x16_c
 
 uint32_t aom_highbd_8_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance16x32 aom_highbd_8_sub_pixel_variance16x32_c
 
+uint32_t aom_highbd_8_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x4 aom_highbd_8_sub_pixel_variance16x4_c
+
 uint32_t aom_highbd_8_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance16x8 aom_highbd_8_sub_pixel_variance16x8_c
 
 uint32_t aom_highbd_8_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance32x16 aom_highbd_8_sub_pixel_variance32x16_c
 
 uint32_t aom_highbd_8_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance32x32 aom_highbd_8_sub_pixel_variance32x32_c
 
 uint32_t aom_highbd_8_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance32x64 aom_highbd_8_sub_pixel_variance32x64_c
 
+uint32_t aom_highbd_8_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x8 aom_highbd_8_sub_pixel_variance32x8_c
+
+uint32_t aom_highbd_8_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x16 aom_highbd_8_sub_pixel_variance4x16_c
+
 uint32_t aom_highbd_8_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance4x4 aom_highbd_8_sub_pixel_variance4x4_c
 
 uint32_t aom_highbd_8_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance4x8 aom_highbd_8_sub_pixel_variance4x8_c
 
 uint32_t aom_highbd_8_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance64x32 aom_highbd_8_sub_pixel_variance64x32_c
 
 uint32_t aom_highbd_8_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance64x64 aom_highbd_8_sub_pixel_variance64x64_c
 
 uint32_t aom_highbd_8_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance8x16 aom_highbd_8_sub_pixel_variance8x16_c
 
+uint32_t aom_highbd_8_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x32 aom_highbd_8_sub_pixel_variance8x32_c
+
 uint32_t aom_highbd_8_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance8x4 aom_highbd_8_sub_pixel_variance8x4_c
 
 uint32_t aom_highbd_8_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance8x8 aom_highbd_8_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_8_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance16x16 aom_highbd_8_variance16x16_c
 
 unsigned int aom_highbd_8_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance16x32 aom_highbd_8_variance16x32_c
 
+unsigned int aom_highbd_8_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x4 aom_highbd_8_variance16x4_c
+
 unsigned int aom_highbd_8_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance16x8 aom_highbd_8_variance16x8_c
 
 unsigned int aom_highbd_8_variance2x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance2x2 aom_highbd_8_variance2x2_c
 
 unsigned int aom_highbd_8_variance2x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance2x4 aom_highbd_8_variance2x4_c
@@ -951,16 +1341,22 @@ unsigned int aom_highbd_8_variance32x16_
 #define aom_highbd_8_variance32x16 aom_highbd_8_variance32x16_c
 
 unsigned int aom_highbd_8_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance32x32 aom_highbd_8_variance32x32_c
 
 unsigned int aom_highbd_8_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance32x64 aom_highbd_8_variance32x64_c
 
+unsigned int aom_highbd_8_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x8 aom_highbd_8_variance32x8_c
+
+unsigned int aom_highbd_8_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x16 aom_highbd_8_variance4x16_c
+
 unsigned int aom_highbd_8_variance4x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance4x2 aom_highbd_8_variance4x2_c
 
 unsigned int aom_highbd_8_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance4x4 aom_highbd_8_variance4x4_c
 
 unsigned int aom_highbd_8_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance4x8 aom_highbd_8_variance4x8_c
@@ -969,16 +1365,19 @@ unsigned int aom_highbd_8_variance64x32_
 #define aom_highbd_8_variance64x32 aom_highbd_8_variance64x32_c
 
 unsigned int aom_highbd_8_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance64x64 aom_highbd_8_variance64x64_c
 
 unsigned int aom_highbd_8_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance8x16 aom_highbd_8_variance8x16_c
 
+unsigned int aom_highbd_8_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x32 aom_highbd_8_variance8x32_c
+
 unsigned int aom_highbd_8_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance8x4 aom_highbd_8_variance8x4_c
 
 unsigned int aom_highbd_8_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance8x8 aom_highbd_8_variance8x8_c
 
 void aom_highbd_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd);
 #define aom_highbd_blend_a64_hmask aom_highbd_blend_a64_hmask_c
@@ -987,23 +1386,23 @@ void aom_highbd_blend_a64_mask_c(uint8_t
 #define aom_highbd_blend_a64_mask aom_highbd_blend_a64_mask_c
 
 void aom_highbd_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd);
 #define aom_highbd_blend_a64_vmask aom_highbd_blend_a64_vmask_c
 
 void aom_highbd_comp_avg_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
 #define aom_highbd_comp_avg_pred aom_highbd_comp_avg_pred_c
 
-void aom_highbd_comp_avg_upsampled_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
+void aom_highbd_comp_avg_upsampled_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref8, int ref_stride, int bd);
 #define aom_highbd_comp_avg_upsampled_pred aom_highbd_comp_avg_upsampled_pred_c
 
 void aom_highbd_comp_mask_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 #define aom_highbd_comp_mask_pred aom_highbd_comp_mask_pred_c
 
-void aom_highbd_comp_mask_upsampled_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_highbd_comp_mask_upsampled_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask, int bd);
 #define aom_highbd_comp_mask_upsampled_pred aom_highbd_comp_mask_upsampled_pred_c
 
 void aom_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
 #define aom_highbd_convolve8 aom_highbd_convolve8_c
 
 void aom_highbd_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
 #define aom_highbd_convolve8_avg aom_highbd_convolve8_avg_c
 
@@ -1023,223 +1422,391 @@ void aom_highbd_convolve_avg_c(const uin
 #define aom_highbd_convolve_avg aom_highbd_convolve_avg_c
 
 void aom_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
 #define aom_highbd_convolve_copy aom_highbd_convolve_copy_c
 
 void aom_highbd_d117_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_16x16 aom_highbd_d117_predictor_16x16_c
 
+void aom_highbd_d117_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_16x32 aom_highbd_d117_predictor_16x32_c
+
+void aom_highbd_d117_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_16x8 aom_highbd_d117_predictor_16x8_c
+
 void aom_highbd_d117_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_2x2 aom_highbd_d117_predictor_2x2_c
 
+void aom_highbd_d117_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_32x16 aom_highbd_d117_predictor_32x16_c
+
 void aom_highbd_d117_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_32x32 aom_highbd_d117_predictor_32x32_c
 
 void aom_highbd_d117_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_4x4 aom_highbd_d117_predictor_4x4_c
 
+void aom_highbd_d117_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_4x8 aom_highbd_d117_predictor_4x8_c
+
+void aom_highbd_d117_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_8x16 aom_highbd_d117_predictor_8x16_c
+
+void aom_highbd_d117_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_8x4 aom_highbd_d117_predictor_8x4_c
+
 void aom_highbd_d117_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_8x8 aom_highbd_d117_predictor_8x8_c
 
 void aom_highbd_d135_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_16x16 aom_highbd_d135_predictor_16x16_c
 
+void aom_highbd_d135_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_16x32 aom_highbd_d135_predictor_16x32_c
+
+void aom_highbd_d135_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_16x8 aom_highbd_d135_predictor_16x8_c
+
 void aom_highbd_d135_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_2x2 aom_highbd_d135_predictor_2x2_c
 
+void aom_highbd_d135_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_32x16 aom_highbd_d135_predictor_32x16_c
+
 void aom_highbd_d135_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_32x32 aom_highbd_d135_predictor_32x32_c
 
 void aom_highbd_d135_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_4x4 aom_highbd_d135_predictor_4x4_c
 
+void aom_highbd_d135_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_4x8 aom_highbd_d135_predictor_4x8_c
+
+void aom_highbd_d135_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_8x16 aom_highbd_d135_predictor_8x16_c
+
+void aom_highbd_d135_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_8x4 aom_highbd_d135_predictor_8x4_c
+
 void aom_highbd_d135_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_8x8 aom_highbd_d135_predictor_8x8_c
 
 void aom_highbd_d153_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_16x16 aom_highbd_d153_predictor_16x16_c
 
+void aom_highbd_d153_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_16x32 aom_highbd_d153_predictor_16x32_c
+
+void aom_highbd_d153_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_16x8 aom_highbd_d153_predictor_16x8_c
+
 void aom_highbd_d153_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_2x2 aom_highbd_d153_predictor_2x2_c
 
+void aom_highbd_d153_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_32x16 aom_highbd_d153_predictor_32x16_c
+
 void aom_highbd_d153_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_32x32 aom_highbd_d153_predictor_32x32_c
 
 void aom_highbd_d153_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_4x4 aom_highbd_d153_predictor_4x4_c
 
+void aom_highbd_d153_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_4x8 aom_highbd_d153_predictor_4x8_c
+
+void aom_highbd_d153_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_8x16 aom_highbd_d153_predictor_8x16_c
+
+void aom_highbd_d153_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_8x4 aom_highbd_d153_predictor_8x4_c
+
 void aom_highbd_d153_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_8x8 aom_highbd_d153_predictor_8x8_c
 
 void aom_highbd_d207e_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_16x16 aom_highbd_d207e_predictor_16x16_c
 
+void aom_highbd_d207e_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_16x32 aom_highbd_d207e_predictor_16x32_c
+
+void aom_highbd_d207e_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_16x8 aom_highbd_d207e_predictor_16x8_c
+
 void aom_highbd_d207e_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_2x2 aom_highbd_d207e_predictor_2x2_c
 
+void aom_highbd_d207e_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_32x16 aom_highbd_d207e_predictor_32x16_c
+
 void aom_highbd_d207e_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_32x32 aom_highbd_d207e_predictor_32x32_c
 
 void aom_highbd_d207e_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_4x4 aom_highbd_d207e_predictor_4x4_c
 
+void aom_highbd_d207e_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_4x8 aom_highbd_d207e_predictor_4x8_c
+
+void aom_highbd_d207e_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_8x16 aom_highbd_d207e_predictor_8x16_c
+
+void aom_highbd_d207e_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_8x4 aom_highbd_d207e_predictor_8x4_c
+
 void aom_highbd_d207e_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_8x8 aom_highbd_d207e_predictor_8x8_c
 
 void aom_highbd_d45e_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_16x16 aom_highbd_d45e_predictor_16x16_c
 
+void aom_highbd_d45e_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_16x32 aom_highbd_d45e_predictor_16x32_c
+
+void aom_highbd_d45e_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_16x8 aom_highbd_d45e_predictor_16x8_c
+
 void aom_highbd_d45e_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_2x2 aom_highbd_d45e_predictor_2x2_c
 
+void aom_highbd_d45e_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_32x16 aom_highbd_d45e_predictor_32x16_c
+
 void aom_highbd_d45e_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_32x32 aom_highbd_d45e_predictor_32x32_c
 
 void aom_highbd_d45e_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_4x4 aom_highbd_d45e_predictor_4x4_c
 
+void aom_highbd_d45e_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_4x8 aom_highbd_d45e_predictor_4x8_c
+
+void aom_highbd_d45e_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_8x16 aom_highbd_d45e_predictor_8x16_c
+
+void aom_highbd_d45e_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_8x4 aom_highbd_d45e_predictor_8x4_c
+
 void aom_highbd_d45e_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_8x8 aom_highbd_d45e_predictor_8x8_c
 
 void aom_highbd_d63e_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_16x16 aom_highbd_d63e_predictor_16x16_c
 
+void aom_highbd_d63e_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_16x32 aom_highbd_d63e_predictor_16x32_c
+
+void aom_highbd_d63e_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_16x8 aom_highbd_d63e_predictor_16x8_c
+
 void aom_highbd_d63e_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_2x2 aom_highbd_d63e_predictor_2x2_c
 
+void aom_highbd_d63e_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_32x16 aom_highbd_d63e_predictor_32x16_c
+
 void aom_highbd_d63e_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_32x32 aom_highbd_d63e_predictor_32x32_c
 
 void aom_highbd_d63e_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_4x4 aom_highbd_d63e_predictor_4x4_c
 
+void aom_highbd_d63e_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_4x8 aom_highbd_d63e_predictor_4x8_c
+
+void aom_highbd_d63e_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_8x16 aom_highbd_d63e_predictor_8x16_c
+
+void aom_highbd_d63e_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_8x4 aom_highbd_d63e_predictor_8x4_c
+
 void aom_highbd_d63e_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_8x8 aom_highbd_d63e_predictor_8x8_c
 
 void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_16x16 aom_highbd_dc_128_predictor_16x16_c
 
+void aom_highbd_dc_128_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x32 aom_highbd_dc_128_predictor_16x32_c
+
+void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_c
+
 void aom_highbd_dc_128_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_c
 
+void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_c
+
 void aom_highbd_dc_128_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_32x32 aom_highbd_dc_128_predictor_32x32_c
 
 void aom_highbd_dc_128_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_4x4 aom_highbd_dc_128_predictor_4x4_c
 
+void aom_highbd_dc_128_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x8 aom_highbd_dc_128_predictor_4x8_c
+
+void aom_highbd_dc_128_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x16 aom_highbd_dc_128_predictor_8x16_c
+
+void aom_highbd_dc_128_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x4 aom_highbd_dc_128_predictor_8x4_c
+
 void aom_highbd_dc_128_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_8x8 aom_highbd_dc_128_predictor_8x8_c
 
 void aom_highbd_dc_left_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_16x16 aom_highbd_dc_left_predictor_16x16_c
 
+void aom_highbd_dc_left_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x32 aom_highbd_dc_left_predictor_16x32_c
+
+void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_c
+
 void aom_highbd_dc_left_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_c
 
+void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_c
+
 void aom_highbd_dc_left_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_32x32 aom_highbd_dc_left_predictor_32x32_c
 
 void aom_highbd_dc_left_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_4x4 aom_highbd_dc_left_predictor_4x4_c
 
+void aom_highbd_dc_left_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x8 aom_highbd_dc_left_predictor_4x8_c
+
+void aom_highbd_dc_left_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x16 aom_highbd_dc_left_predictor_8x16_c
+
+void aom_highbd_dc_left_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x4 aom_highbd_dc_left_predictor_8x4_c
+
 void aom_highbd_dc_left_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_8x8 aom_highbd_dc_left_predictor_8x8_c
 
 void aom_highbd_dc_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_16x16 aom_highbd_dc_predictor_16x16_c
 
+void aom_highbd_dc_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x32 aom_highbd_dc_predictor_16x32_c
+
+void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_c
+
 void aom_highbd_dc_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_c
 
+void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_c
+
 void aom_highbd_dc_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_32x32 aom_highbd_dc_predictor_32x32_c
 
 void aom_highbd_dc_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_4x4 aom_highbd_dc_predictor_4x4_c
 
+void aom_highbd_dc_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x8 aom_highbd_dc_predictor_4x8_c
+
+void aom_highbd_dc_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x16 aom_highbd_dc_predictor_8x16_c
+
+void aom_highbd_dc_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x4 aom_highbd_dc_predictor_8x4_c
+
 void aom_highbd_dc_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_8x8 aom_highbd_dc_predictor_8x8_c
 
 void aom_highbd_dc_top_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_16x16 aom_highbd_dc_top_predictor_16x16_c
 
+void aom_highbd_dc_top_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x32 aom_highbd_dc_top_predictor_16x32_c
+
+void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_c
+
 void aom_highbd_dc_top_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_c
 
+void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_c
+
 void aom_highbd_dc_top_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_32x32 aom_highbd_dc_top_predictor_32x32_c
 
 void aom_highbd_dc_top_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_4x4 aom_highbd_dc_top_predictor_4x4_c
 
+void aom_highbd_dc_top_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x8 aom_highbd_dc_top_predictor_4x8_c
+
+void aom_highbd_dc_top_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x16 aom_highbd_dc_top_predictor_8x16_c
+
+void aom_highbd_dc_top_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x4 aom_highbd_dc_top_predictor_8x4_c
+
 void aom_highbd_dc_top_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_8x8 aom_highbd_dc_top_predictor_8x8_c
 
 void aom_highbd_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_highbd_fdct16x16 aom_highbd_fdct16x16_c
 
-void aom_highbd_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
-#define aom_highbd_fdct16x16_1 aom_highbd_fdct16x16_1_c
-
 void aom_highbd_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_highbd_fdct32x32 aom_highbd_fdct32x32_c
 
-void aom_highbd_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
-#define aom_highbd_fdct32x32_1 aom_highbd_fdct32x32_1_c
-
 void aom_highbd_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_highbd_fdct32x32_rd aom_highbd_fdct32x32_rd_c
 
 void aom_highbd_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_highbd_fdct4x4 aom_highbd_fdct4x4_c
 
 void aom_highbd_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_highbd_fdct8x8 aom_highbd_fdct8x8_c
 
-void aom_highbd_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
-#define aom_highbd_fdct8x8_1 aom_highbd_fdct8x8_1_c
-
 void aom_highbd_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_h_predictor_16x16 aom_highbd_h_predictor_16x16_c
 
+void aom_highbd_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x32 aom_highbd_h_predictor_16x32_c
+
+void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x8 aom_highbd_h_predictor_16x8_c
+
 void aom_highbd_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_h_predictor_2x2 aom_highbd_h_predictor_2x2_c
 
+void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x16 aom_highbd_h_predictor_32x16_c
+
 void aom_highbd_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_h_predictor_32x32 aom_highbd_h_predictor_32x32_c
 
 void aom_highbd_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_h_predictor_4x4 aom_highbd_h_predictor_4x4_c
 
+void aom_highbd_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x8 aom_highbd_h_predictor_4x8_c
+
+void aom_highbd_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x16 aom_highbd_h_predictor_8x16_c
+
+void aom_highbd_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x4 aom_highbd_h_predictor_8x4_c
+
 void aom_highbd_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_h_predictor_8x8 aom_highbd_h_predictor_8x8_c
 
-void aom_highbd_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-#define aom_highbd_idct16x16_1_add aom_highbd_idct16x16_1_add_c
-
-void aom_highbd_idct32x32_1024_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-#define aom_highbd_idct32x32_1024_add aom_highbd_idct32x32_1024_add_c
-
-void aom_highbd_idct32x32_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-#define aom_highbd_idct32x32_1_add aom_highbd_idct32x32_1_add_c
-
-void aom_highbd_idct32x32_34_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-#define aom_highbd_idct32x32_34_add aom_highbd_idct32x32_34_add_c
-
-void aom_highbd_idct4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-#define aom_highbd_idct4x4_16_add aom_highbd_idct4x4_16_add_c
-
-void aom_highbd_idct4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-#define aom_highbd_idct4x4_1_add aom_highbd_idct4x4_1_add_c
-
-void aom_highbd_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
-#define aom_highbd_idct8x8_1_add aom_highbd_idct8x8_1_add_c
-
 void aom_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
 #define aom_highbd_iwht4x4_16_add aom_highbd_iwht4x4_16_add_c
 
 void aom_highbd_iwht4x4_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
 #define aom_highbd_iwht4x4_1_add aom_highbd_iwht4x4_1_add_c
 
 void aom_highbd_lpf_horizontal_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
 #define aom_highbd_lpf_horizontal_4 aom_highbd_lpf_horizontal_4_c
@@ -1278,181 +1845,247 @@ void aom_highbd_lpf_vertical_8_dual_c(ui
 #define aom_highbd_lpf_vertical_8_dual aom_highbd_lpf_vertical_8_dual_c
 
 unsigned int aom_highbd_masked_sad16x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_highbd_masked_sad16x16 aom_highbd_masked_sad16x16_c
 
 unsigned int aom_highbd_masked_sad16x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_highbd_masked_sad16x32 aom_highbd_masked_sad16x32_c
 
+unsigned int aom_highbd_masked_sad16x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad16x4 aom_highbd_masked_sad16x4_c
+
 unsigned int aom_highbd_masked_sad16x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_highbd_masked_sad16x8 aom_highbd_masked_sad16x8_c
 
 unsigned int aom_highbd_masked_sad32x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_highbd_masked_sad32x16 aom_highbd_masked_sad32x16_c
 
 unsigned int aom_highbd_masked_sad32x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_highbd_masked_sad32x32 aom_highbd_masked_sad32x32_c
 
 unsigned int aom_highbd_masked_sad32x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_highbd_masked_sad32x64 aom_highbd_masked_sad32x64_c
 
+unsigned int aom_highbd_masked_sad32x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad32x8 aom_highbd_masked_sad32x8_c
+
+unsigned int aom_highbd_masked_sad4x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad4x16 aom_highbd_masked_sad4x16_c
+
 unsigned int aom_highbd_masked_sad4x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_highbd_masked_sad4x4 aom_highbd_masked_sad4x4_c
 
 unsigned int aom_highbd_masked_sad4x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_highbd_masked_sad4x8 aom_highbd_masked_sad4x8_c
 
 unsigned int aom_highbd_masked_sad64x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_highbd_masked_sad64x32 aom_highbd_masked_sad64x32_c
 
 unsigned int aom_highbd_masked_sad64x64_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_highbd_masked_sad64x64 aom_highbd_masked_sad64x64_c
 
 unsigned int aom_highbd_masked_sad8x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_highbd_masked_sad8x16 aom_highbd_masked_sad8x16_c
 
+unsigned int aom_highbd_masked_sad8x32_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_highbd_masked_sad8x32 aom_highbd_masked_sad8x32_c
+
 unsigned int aom_highbd_masked_sad8x4_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_highbd_masked_sad8x4 aom_highbd_masked_sad8x4_c
 
 unsigned int aom_highbd_masked_sad8x8_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_highbd_masked_sad8x8 aom_highbd_masked_sad8x8_c
 
 void aom_highbd_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
 #define aom_highbd_minmax_8x8 aom_highbd_minmax_8x8_c
 
 unsigned int aom_highbd_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_highbd_obmc_sad16x16 aom_highbd_obmc_sad16x16_c
 
 unsigned int aom_highbd_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_highbd_obmc_sad16x32 aom_highbd_obmc_sad16x32_c
 
+unsigned int aom_highbd_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad16x4 aom_highbd_obmc_sad16x4_c
+
 unsigned int aom_highbd_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_highbd_obmc_sad16x8 aom_highbd_obmc_sad16x8_c
 
 unsigned int aom_highbd_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_highbd_obmc_sad32x16 aom_highbd_obmc_sad32x16_c
 
 unsigned int aom_highbd_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_highbd_obmc_sad32x32 aom_highbd_obmc_sad32x32_c
 
 unsigned int aom_highbd_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_highbd_obmc_sad32x64 aom_highbd_obmc_sad32x64_c
 
+unsigned int aom_highbd_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad32x8 aom_highbd_obmc_sad32x8_c
+
+unsigned int aom_highbd_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad4x16 aom_highbd_obmc_sad4x16_c
+
 unsigned int aom_highbd_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_highbd_obmc_sad4x4 aom_highbd_obmc_sad4x4_c
 
 unsigned int aom_highbd_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_highbd_obmc_sad4x8 aom_highbd_obmc_sad4x8_c
 
 unsigned int aom_highbd_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_highbd_obmc_sad64x32 aom_highbd_obmc_sad64x32_c
 
 unsigned int aom_highbd_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_highbd_obmc_sad64x64 aom_highbd_obmc_sad64x64_c
 
 unsigned int aom_highbd_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_highbd_obmc_sad8x16 aom_highbd_obmc_sad8x16_c
 
+unsigned int aom_highbd_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_highbd_obmc_sad8x32 aom_highbd_obmc_sad8x32_c
+
 unsigned int aom_highbd_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_highbd_obmc_sad8x4 aom_highbd_obmc_sad8x4_c
 
 unsigned int aom_highbd_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_highbd_obmc_sad8x8 aom_highbd_obmc_sad8x8_c
 
 unsigned int aom_highbd_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_sub_pixel_variance16x16 aom_highbd_obmc_sub_pixel_variance16x16_c
 
 unsigned int aom_highbd_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_sub_pixel_variance16x32 aom_highbd_obmc_sub_pixel_variance16x32_c
 
+unsigned int aom_highbd_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance16x4 aom_highbd_obmc_sub_pixel_variance16x4_c
+
 unsigned int aom_highbd_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_sub_pixel_variance16x8 aom_highbd_obmc_sub_pixel_variance16x8_c
 
 unsigned int aom_highbd_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_sub_pixel_variance32x16 aom_highbd_obmc_sub_pixel_variance32x16_c
 
 unsigned int aom_highbd_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_sub_pixel_variance32x32 aom_highbd_obmc_sub_pixel_variance32x32_c
 
 unsigned int aom_highbd_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_sub_pixel_variance32x64 aom_highbd_obmc_sub_pixel_variance32x64_c
 
+unsigned int aom_highbd_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance32x8 aom_highbd_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance4x16 aom_highbd_obmc_sub_pixel_variance4x16_c
+
 unsigned int aom_highbd_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_sub_pixel_variance4x4 aom_highbd_obmc_sub_pixel_variance4x4_c
 
 unsigned int aom_highbd_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_sub_pixel_variance4x8 aom_highbd_obmc_sub_pixel_variance4x8_c
 
 unsigned int aom_highbd_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_sub_pixel_variance64x32 aom_highbd_obmc_sub_pixel_variance64x32_c
 
 unsigned int aom_highbd_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_sub_pixel_variance64x64 aom_highbd_obmc_sub_pixel_variance64x64_c
 
 unsigned int aom_highbd_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_sub_pixel_variance8x16 aom_highbd_obmc_sub_pixel_variance8x16_c
 
+unsigned int aom_highbd_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_sub_pixel_variance8x32 aom_highbd_obmc_sub_pixel_variance8x32_c
+
 unsigned int aom_highbd_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_sub_pixel_variance8x4 aom_highbd_obmc_sub_pixel_variance8x4_c
 
 unsigned int aom_highbd_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_sub_pixel_variance8x8 aom_highbd_obmc_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_variance16x16 aom_highbd_obmc_variance16x16_c
 
 unsigned int aom_highbd_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_variance16x32 aom_highbd_obmc_variance16x32_c
 
+unsigned int aom_highbd_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_variance16x4 aom_highbd_obmc_variance16x4_c
+
 unsigned int aom_highbd_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_variance16x8 aom_highbd_obmc_variance16x8_c
 
 unsigned int aom_highbd_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_variance32x16 aom_highbd_obmc_variance32x16_c
 
 unsigned int aom_highbd_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_variance32x32 aom_highbd_obmc_variance32x32_c
 
 unsigned int aom_highbd_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_variance32x64 aom_highbd_obmc_variance32x64_c
 
+unsigned int aom_highbd_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_variance32x8 aom_highbd_obmc_variance32x8_c
+
+unsigned int aom_highbd_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_variance4x16 aom_highbd_obmc_variance4x16_c
+
 unsigned int aom_highbd_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_variance4x4 aom_highbd_obmc_variance4x4_c
 
 unsigned int aom_highbd_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_variance4x8 aom_highbd_obmc_variance4x8_c
 
 unsigned int aom_highbd_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_variance64x32 aom_highbd_obmc_variance64x32_c
 
 unsigned int aom_highbd_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_variance64x64 aom_highbd_obmc_variance64x64_c
 
 unsigned int aom_highbd_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_variance8x16 aom_highbd_obmc_variance8x16_c
 
+unsigned int aom_highbd_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_obmc_variance8x32 aom_highbd_obmc_variance8x32_c
+
 unsigned int aom_highbd_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_variance8x4 aom_highbd_obmc_variance8x4_c
 
 unsigned int aom_highbd_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_obmc_variance8x8 aom_highbd_obmc_variance8x8_c
 
 void aom_highbd_paeth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_paeth_predictor_16x16 aom_highbd_paeth_predictor_16x16_c
 
+void aom_highbd_paeth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x32 aom_highbd_paeth_predictor_16x32_c
+
+void aom_highbd_paeth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_16x8 aom_highbd_paeth_predictor_16x8_c
+
 void aom_highbd_paeth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_paeth_predictor_2x2 aom_highbd_paeth_predictor_2x2_c
 
+void aom_highbd_paeth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_32x16 aom_highbd_paeth_predictor_32x16_c
+
 void aom_highbd_paeth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_paeth_predictor_32x32 aom_highbd_paeth_predictor_32x32_c
 
 void aom_highbd_paeth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_paeth_predictor_4x4 aom_highbd_paeth_predictor_4x4_c
 
+void aom_highbd_paeth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_4x8 aom_highbd_paeth_predictor_4x8_c
+
+void aom_highbd_paeth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x16 aom_highbd_paeth_predictor_8x16_c
+
+void aom_highbd_paeth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_8x4 aom_highbd_paeth_predictor_8x4_c
+
 void aom_highbd_paeth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_paeth_predictor_8x8 aom_highbd_paeth_predictor_8x8_c
 
 void aom_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 #define aom_highbd_quantize_b aom_highbd_quantize_b_c
 
 void aom_highbd_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 #define aom_highbd_quantize_b_32x32 aom_highbd_quantize_b_32x32_c
@@ -1479,16 +2112,25 @@ unsigned int aom_highbd_sad16x32_c(const
 #define aom_highbd_sad16x32 aom_highbd_sad16x32_c
 
 unsigned int aom_highbd_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define aom_highbd_sad16x32_avg aom_highbd_sad16x32_avg_c
 
 void aom_highbd_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
 #define aom_highbd_sad16x32x4d aom_highbd_sad16x32x4d_c
 
+unsigned int aom_highbd_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad16x4 aom_highbd_sad16x4_c
+
+unsigned int aom_highbd_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad16x4_avg aom_highbd_sad16x4_avg_c
+
+void aom_highbd_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad16x4x4d aom_highbd_sad16x4x4d_c
+
 unsigned int aom_highbd_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
 #define aom_highbd_sad16x8 aom_highbd_sad16x8_c
 
 unsigned int aom_highbd_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define aom_highbd_sad16x8_avg aom_highbd_sad16x8_avg_c
 
 void aom_highbd_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define aom_highbd_sad16x8x3 aom_highbd_sad16x8x3_c
@@ -1527,16 +2169,34 @@ unsigned int aom_highbd_sad32x64_c(const
 #define aom_highbd_sad32x64 aom_highbd_sad32x64_c
 
 unsigned int aom_highbd_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define aom_highbd_sad32x64_avg aom_highbd_sad32x64_avg_c
 
 void aom_highbd_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
 #define aom_highbd_sad32x64x4d aom_highbd_sad32x64x4d_c
 
+unsigned int aom_highbd_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad32x8 aom_highbd_sad32x8_c
+
+unsigned int aom_highbd_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad32x8_avg aom_highbd_sad32x8_avg_c
+
+void aom_highbd_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad32x8x4d aom_highbd_sad32x8x4d_c
+
+unsigned int aom_highbd_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad4x16 aom_highbd_sad4x16_c
+
+unsigned int aom_highbd_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad4x16_avg aom_highbd_sad4x16_avg_c
+
+void aom_highbd_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad4x16x4d aom_highbd_sad4x16x4d_c
+
 unsigned int aom_highbd_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
 #define aom_highbd_sad4x4 aom_highbd_sad4x4_c
 
 unsigned int aom_highbd_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define aom_highbd_sad4x4_avg aom_highbd_sad4x4_avg_c
 
 void aom_highbd_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define aom_highbd_sad4x4x3 aom_highbd_sad4x4x3_c
@@ -1593,16 +2253,25 @@ void aom_highbd_sad8x16x3_c(const uint8_
 #define aom_highbd_sad8x16x3 aom_highbd_sad8x16x3_c
 
 void aom_highbd_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
 #define aom_highbd_sad8x16x4d aom_highbd_sad8x16x4d_c
 
 void aom_highbd_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define aom_highbd_sad8x16x8 aom_highbd_sad8x16x8_c
 
+unsigned int aom_highbd_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_highbd_sad8x32 aom_highbd_sad8x32_c
+
+unsigned int aom_highbd_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_highbd_sad8x32_avg aom_highbd_sad8x32_avg_c
+
+void aom_highbd_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_highbd_sad8x32x4d aom_highbd_sad8x32x4d_c
+
 unsigned int aom_highbd_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
 #define aom_highbd_sad8x4 aom_highbd_sad8x4_c
 
 unsigned int aom_highbd_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define aom_highbd_sad8x4_avg aom_highbd_sad8x4_avg_c
 
 void aom_highbd_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
 #define aom_highbd_sad8x4x4d aom_highbd_sad8x4x4d_c
@@ -1620,49 +2289,151 @@ void aom_highbd_sad8x8x3_c(const uint8_t
 #define aom_highbd_sad8x8x3 aom_highbd_sad8x8x3_c
 
 void aom_highbd_sad8x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
 #define aom_highbd_sad8x8x4d aom_highbd_sad8x8x4d_c
 
 void aom_highbd_sad8x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define aom_highbd_sad8x8x8 aom_highbd_sad8x8x8_c
 
+void aom_highbd_smooth_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x16 aom_highbd_smooth_h_predictor_16x16_c
+
+void aom_highbd_smooth_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x32 aom_highbd_smooth_h_predictor_16x32_c
+
+void aom_highbd_smooth_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_16x8 aom_highbd_smooth_h_predictor_16x8_c
+
+void aom_highbd_smooth_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_2x2 aom_highbd_smooth_h_predictor_2x2_c
+
+void aom_highbd_smooth_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x16 aom_highbd_smooth_h_predictor_32x16_c
+
+void aom_highbd_smooth_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_32x32 aom_highbd_smooth_h_predictor_32x32_c
+
+void aom_highbd_smooth_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x4 aom_highbd_smooth_h_predictor_4x4_c
+
+void aom_highbd_smooth_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_4x8 aom_highbd_smooth_h_predictor_4x8_c
+
+void aom_highbd_smooth_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x16 aom_highbd_smooth_h_predictor_8x16_c
+
+void aom_highbd_smooth_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x4 aom_highbd_smooth_h_predictor_8x4_c
+
+void aom_highbd_smooth_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_8x8 aom_highbd_smooth_h_predictor_8x8_c
+
 void aom_highbd_smooth_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_smooth_predictor_16x16 aom_highbd_smooth_predictor_16x16_c
 
+void aom_highbd_smooth_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x32 aom_highbd_smooth_predictor_16x32_c
+
+void aom_highbd_smooth_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_16x8 aom_highbd_smooth_predictor_16x8_c
+
 void aom_highbd_smooth_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_smooth_predictor_2x2 aom_highbd_smooth_predictor_2x2_c
 
+void aom_highbd_smooth_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_32x16 aom_highbd_smooth_predictor_32x16_c
+
 void aom_highbd_smooth_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_smooth_predictor_32x32 aom_highbd_smooth_predictor_32x32_c
 
 void aom_highbd_smooth_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_smooth_predictor_4x4 aom_highbd_smooth_predictor_4x4_c
 
+void aom_highbd_smooth_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_4x8 aom_highbd_smooth_predictor_4x8_c
+
+void aom_highbd_smooth_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x16 aom_highbd_smooth_predictor_8x16_c
+
+void aom_highbd_smooth_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_8x4 aom_highbd_smooth_predictor_8x4_c
+
 void aom_highbd_smooth_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_smooth_predictor_8x8 aom_highbd_smooth_predictor_8x8_c
 
+void aom_highbd_smooth_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x16 aom_highbd_smooth_v_predictor_16x16_c
+
+void aom_highbd_smooth_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x32 aom_highbd_smooth_v_predictor_16x32_c
+
+void aom_highbd_smooth_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_16x8 aom_highbd_smooth_v_predictor_16x8_c
+
+void aom_highbd_smooth_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_2x2 aom_highbd_smooth_v_predictor_2x2_c
+
+void aom_highbd_smooth_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x16 aom_highbd_smooth_v_predictor_32x16_c
+
+void aom_highbd_smooth_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_32x32 aom_highbd_smooth_v_predictor_32x32_c
+
+void aom_highbd_smooth_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x4 aom_highbd_smooth_v_predictor_4x4_c
+
+void aom_highbd_smooth_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_4x8 aom_highbd_smooth_v_predictor_4x8_c
+
+void aom_highbd_smooth_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x16 aom_highbd_smooth_v_predictor_8x16_c
+
+void aom_highbd_smooth_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x4 aom_highbd_smooth_v_predictor_8x4_c
+
+void aom_highbd_smooth_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_8x8 aom_highbd_smooth_v_predictor_8x8_c
+
 void aom_highbd_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride, int bd);
 #define aom_highbd_subtract_block aom_highbd_subtract_block_c
 
-void aom_highbd_upsampled_pred_c(uint16_t *comp_pred, int width, int height, const uint8_t *ref8, int ref_stride);
+void aom_highbd_upsampled_pred_c(uint16_t *comp_pred, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref8, int ref_stride, int bd);
 #define aom_highbd_upsampled_pred aom_highbd_upsampled_pred_c
 
 void aom_highbd_v_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_v_predictor_16x16 aom_highbd_v_predictor_16x16_c
 
+void aom_highbd_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x32 aom_highbd_v_predictor_16x32_c
+
+void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x8 aom_highbd_v_predictor_16x8_c
+
 void aom_highbd_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_v_predictor_2x2 aom_highbd_v_predictor_2x2_c
 
+void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x16 aom_highbd_v_predictor_32x16_c
+
 void aom_highbd_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_v_predictor_32x32 aom_highbd_v_predictor_32x32_c
 
 void aom_highbd_v_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_v_predictor_4x4 aom_highbd_v_predictor_4x4_c
 
+void aom_highbd_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x8 aom_highbd_v_predictor_4x8_c
+
+void aom_highbd_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x16 aom_highbd_v_predictor_8x16_c
+
+void aom_highbd_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x4 aom_highbd_v_predictor_8x4_c
+
 void aom_highbd_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_v_predictor_8x8 aom_highbd_v_predictor_8x8_c
 
 void aom_idct16x16_10_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
 #define aom_idct16x16_10_add aom_idct16x16_10_add_c
 
 void aom_idct16x16_1_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride);
 #define aom_idct16x16_1_add aom_idct16x16_1_add_c
@@ -1749,82 +2520,106 @@ void aom_lpf_vertical_8_dual_c(uint8_t *
 #define aom_lpf_vertical_8_dual aom_lpf_vertical_8_dual_c
 
 unsigned int aom_masked_sad16x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_masked_sad16x16 aom_masked_sad16x16_c
 
 unsigned int aom_masked_sad16x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_masked_sad16x32 aom_masked_sad16x32_c
 
+unsigned int aom_masked_sad16x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad16x4 aom_masked_sad16x4_c
+
 unsigned int aom_masked_sad16x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_masked_sad16x8 aom_masked_sad16x8_c
 
 unsigned int aom_masked_sad32x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_masked_sad32x16 aom_masked_sad32x16_c
 
 unsigned int aom_masked_sad32x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_masked_sad32x32 aom_masked_sad32x32_c
 
 unsigned int aom_masked_sad32x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_masked_sad32x64 aom_masked_sad32x64_c
 
+unsigned int aom_masked_sad32x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad32x8 aom_masked_sad32x8_c
+
+unsigned int aom_masked_sad4x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad4x16 aom_masked_sad4x16_c
+
 unsigned int aom_masked_sad4x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_masked_sad4x4 aom_masked_sad4x4_c
 
 unsigned int aom_masked_sad4x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_masked_sad4x8 aom_masked_sad4x8_c
 
 unsigned int aom_masked_sad64x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_masked_sad64x32 aom_masked_sad64x32_c
 
 unsigned int aom_masked_sad64x64_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_masked_sad64x64 aom_masked_sad64x64_c
 
 unsigned int aom_masked_sad8x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_masked_sad8x16 aom_masked_sad8x16_c
 
+unsigned int aom_masked_sad8x32_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
+#define aom_masked_sad8x32 aom_masked_sad8x32_c
+
 unsigned int aom_masked_sad8x4_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_masked_sad8x4 aom_masked_sad8x4_c
 
 unsigned int aom_masked_sad8x8_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
 #define aom_masked_sad8x8 aom_masked_sad8x8_c
 
 unsigned int aom_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_masked_sub_pixel_variance16x16 aom_masked_sub_pixel_variance16x16_c
 
 unsigned int aom_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_masked_sub_pixel_variance16x32 aom_masked_sub_pixel_variance16x32_c
 
+unsigned int aom_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance16x4 aom_masked_sub_pixel_variance16x4_c
+
 unsigned int aom_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_masked_sub_pixel_variance16x8 aom_masked_sub_pixel_variance16x8_c
 
 unsigned int aom_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_masked_sub_pixel_variance32x16 aom_masked_sub_pixel_variance32x16_c
 
 unsigned int aom_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_masked_sub_pixel_variance32x32 aom_masked_sub_pixel_variance32x32_c
 
 unsigned int aom_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_masked_sub_pixel_variance32x64 aom_masked_sub_pixel_variance32x64_c
 
+unsigned int aom_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance32x8 aom_masked_sub_pixel_variance32x8_c
+
+unsigned int aom_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance4x16 aom_masked_sub_pixel_variance4x16_c
+
 unsigned int aom_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_masked_sub_pixel_variance4x4 aom_masked_sub_pixel_variance4x4_c
 
 unsigned int aom_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_masked_sub_pixel_variance4x8 aom_masked_sub_pixel_variance4x8_c
 
 unsigned int aom_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_masked_sub_pixel_variance64x32 aom_masked_sub_pixel_variance64x32_c
 
 unsigned int aom_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_masked_sub_pixel_variance64x64 aom_masked_sub_pixel_variance64x64_c
 
 unsigned int aom_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_masked_sub_pixel_variance8x16 aom_masked_sub_pixel_variance8x16_c
 
+unsigned int aom_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_masked_sub_pixel_variance8x32 aom_masked_sub_pixel_variance8x32_c
+
 unsigned int aom_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_masked_sub_pixel_variance8x4 aom_masked_sub_pixel_variance8x4_c
 
 unsigned int aom_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_masked_sub_pixel_variance8x8 aom_masked_sub_pixel_variance8x8_c
 
 void aom_minmax_8x8_c(const uint8_t *s, int p, const uint8_t *d, int dp, int *min, int *max);
 #define aom_minmax_8x8 aom_minmax_8x8_c
@@ -1842,139 +2637,193 @@ unsigned int aom_mse8x8_c(const uint8_t 
 #define aom_mse8x8 aom_mse8x8_c
 
 unsigned int aom_obmc_sad16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_obmc_sad16x16 aom_obmc_sad16x16_c
 
 unsigned int aom_obmc_sad16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_obmc_sad16x32 aom_obmc_sad16x32_c
 
+unsigned int aom_obmc_sad16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad16x4 aom_obmc_sad16x4_c
+
 unsigned int aom_obmc_sad16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_obmc_sad16x8 aom_obmc_sad16x8_c
 
 unsigned int aom_obmc_sad32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_obmc_sad32x16 aom_obmc_sad32x16_c
 
 unsigned int aom_obmc_sad32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_obmc_sad32x32 aom_obmc_sad32x32_c
 
 unsigned int aom_obmc_sad32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_obmc_sad32x64 aom_obmc_sad32x64_c
 
+unsigned int aom_obmc_sad32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad32x8 aom_obmc_sad32x8_c
+
+unsigned int aom_obmc_sad4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad4x16 aom_obmc_sad4x16_c
+
 unsigned int aom_obmc_sad4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_obmc_sad4x4 aom_obmc_sad4x4_c
 
 unsigned int aom_obmc_sad4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_obmc_sad4x8 aom_obmc_sad4x8_c
 
 unsigned int aom_obmc_sad64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_obmc_sad64x32 aom_obmc_sad64x32_c
 
 unsigned int aom_obmc_sad64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_obmc_sad64x64 aom_obmc_sad64x64_c
 
 unsigned int aom_obmc_sad8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_obmc_sad8x16 aom_obmc_sad8x16_c
 
+unsigned int aom_obmc_sad8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
+#define aom_obmc_sad8x32 aom_obmc_sad8x32_c
+
 unsigned int aom_obmc_sad8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_obmc_sad8x4 aom_obmc_sad8x4_c
 
 unsigned int aom_obmc_sad8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask);
 #define aom_obmc_sad8x8 aom_obmc_sad8x8_c
 
 unsigned int aom_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_sub_pixel_variance16x16 aom_obmc_sub_pixel_variance16x16_c
 
 unsigned int aom_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_sub_pixel_variance16x32 aom_obmc_sub_pixel_variance16x32_c
 
+unsigned int aom_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance16x4 aom_obmc_sub_pixel_variance16x4_c
+
 unsigned int aom_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_sub_pixel_variance16x8 aom_obmc_sub_pixel_variance16x8_c
 
 unsigned int aom_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_sub_pixel_variance32x16 aom_obmc_sub_pixel_variance32x16_c
 
 unsigned int aom_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_sub_pixel_variance32x32 aom_obmc_sub_pixel_variance32x32_c
 
 unsigned int aom_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_sub_pixel_variance32x64 aom_obmc_sub_pixel_variance32x64_c
 
+unsigned int aom_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance32x8 aom_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance4x16 aom_obmc_sub_pixel_variance4x16_c
+
 unsigned int aom_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_sub_pixel_variance4x4 aom_obmc_sub_pixel_variance4x4_c
 
 unsigned int aom_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_sub_pixel_variance4x8 aom_obmc_sub_pixel_variance4x8_c
 
 unsigned int aom_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_sub_pixel_variance64x32 aom_obmc_sub_pixel_variance64x32_c
 
 unsigned int aom_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_sub_pixel_variance64x64 aom_obmc_sub_pixel_variance64x64_c
 
 unsigned int aom_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_sub_pixel_variance8x16 aom_obmc_sub_pixel_variance8x16_c
 
+unsigned int aom_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_sub_pixel_variance8x32 aom_obmc_sub_pixel_variance8x32_c
+
 unsigned int aom_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_sub_pixel_variance8x4 aom_obmc_sub_pixel_variance8x4_c
 
 unsigned int aom_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_sub_pixel_variance8x8 aom_obmc_sub_pixel_variance8x8_c
 
 unsigned int aom_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_variance16x16 aom_obmc_variance16x16_c
 
 unsigned int aom_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_variance16x32 aom_obmc_variance16x32_c
 
+unsigned int aom_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance16x4 aom_obmc_variance16x4_c
+
 unsigned int aom_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_variance16x8 aom_obmc_variance16x8_c
 
 unsigned int aom_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_variance32x16 aom_obmc_variance32x16_c
 
 unsigned int aom_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_variance32x32 aom_obmc_variance32x32_c
 
 unsigned int aom_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_variance32x64 aom_obmc_variance32x64_c
 
+unsigned int aom_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance32x8 aom_obmc_variance32x8_c
+
+unsigned int aom_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance4x16 aom_obmc_variance4x16_c
+
 unsigned int aom_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_variance4x4 aom_obmc_variance4x4_c
 
 unsigned int aom_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_variance4x8 aom_obmc_variance4x8_c
 
 unsigned int aom_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_variance64x32 aom_obmc_variance64x32_c
 
 unsigned int aom_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_variance64x64 aom_obmc_variance64x64_c
 
 unsigned int aom_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_variance8x16 aom_obmc_variance8x16_c
 
+unsigned int aom_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_obmc_variance8x32 aom_obmc_variance8x32_c
+
 unsigned int aom_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_variance8x4 aom_obmc_variance8x4_c
 
 unsigned int aom_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_obmc_variance8x8 aom_obmc_variance8x8_c
 
 void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_c
 
+void aom_paeth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_c
+
+void aom_paeth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_c
+
 void aom_paeth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
 
+void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_c
+
 void aom_paeth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_c
 
 void aom_paeth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_c
 
+void aom_paeth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_c
+
+void aom_paeth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_c
+
+void aom_paeth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_c
+
 void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_c
 
 void aom_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 #define aom_quantize_b aom_quantize_b_c
 
 void aom_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 #define aom_quantize_b_32x32 aom_quantize_b_32x32_c
@@ -2001,16 +2850,25 @@ unsigned int aom_sad16x32_c(const uint8_
 #define aom_sad16x32 aom_sad16x32_c
 
 unsigned int aom_sad16x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define aom_sad16x32_avg aom_sad16x32_avg_c
 
 void aom_sad16x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
 #define aom_sad16x32x4d aom_sad16x32x4d_c
 
+unsigned int aom_sad16x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad16x4 aom_sad16x4_c
+
+unsigned int aom_sad16x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad16x4_avg aom_sad16x4_avg_c
+
+void aom_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad16x4x4d aom_sad16x4x4d_c
+
 unsigned int aom_sad16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
 #define aom_sad16x8 aom_sad16x8_c
 
 unsigned int aom_sad16x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define aom_sad16x8_avg aom_sad16x8_avg_c
 
 void aom_sad16x8x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define aom_sad16x8x3 aom_sad16x8x3_c
@@ -2049,16 +2907,34 @@ unsigned int aom_sad32x64_c(const uint8_
 #define aom_sad32x64 aom_sad32x64_c
 
 unsigned int aom_sad32x64_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define aom_sad32x64_avg aom_sad32x64_avg_c
 
 void aom_sad32x64x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
 #define aom_sad32x64x4d aom_sad32x64x4d_c
 
+unsigned int aom_sad32x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad32x8 aom_sad32x8_c
+
+unsigned int aom_sad32x8_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad32x8_avg aom_sad32x8_avg_c
+
+void aom_sad32x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad32x8x4d aom_sad32x8x4d_c
+
+unsigned int aom_sad4x16_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad4x16 aom_sad4x16_c
+
+unsigned int aom_sad4x16_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad4x16_avg aom_sad4x16_avg_c
+
+void aom_sad4x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad4x16x4d aom_sad4x16x4d_c
+
 unsigned int aom_sad4x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
 #define aom_sad4x4 aom_sad4x4_c
 
 unsigned int aom_sad4x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define aom_sad4x4_avg aom_sad4x4_avg_c
 
 void aom_sad4x4x3_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define aom_sad4x4x3 aom_sad4x4x3_c
@@ -2115,16 +2991,25 @@ void aom_sad8x16x3_c(const uint8_t *src_
 #define aom_sad8x16x3 aom_sad8x16x3_c
 
 void aom_sad8x16x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
 #define aom_sad8x16x4d aom_sad8x16x4d_c
 
 void aom_sad8x16x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
 #define aom_sad8x16x8 aom_sad8x16x8_c
 
+unsigned int aom_sad8x32_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
+#define aom_sad8x32 aom_sad8x32_c
+
+unsigned int aom_sad8x32_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
+#define aom_sad8x32_avg aom_sad8x32_avg_c
+
+void aom_sad8x32x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
+#define aom_sad8x32x4d aom_sad8x32x4d_c
+
 unsigned int aom_sad8x4_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
 #define aom_sad8x4 aom_sad8x4_c
 
 unsigned int aom_sad8x4_avg_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 #define aom_sad8x4_avg aom_sad8x4_avg_c
 
 void aom_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array);
 #define aom_sad8x4x4d aom_sad8x4x4d_c
@@ -2163,142 +3048,271 @@ void aom_scaled_avg_vert_c(const uint8_t
 #define aom_scaled_avg_vert aom_scaled_avg_vert_c
 
 void aom_scaled_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 #define aom_scaled_horiz aom_scaled_horiz_c
 
 void aom_scaled_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 #define aom_scaled_vert aom_scaled_vert_c
 
+void aom_smooth_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_16x16 aom_smooth_h_predictor_16x16_c
+
+void aom_smooth_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_16x32 aom_smooth_h_predictor_16x32_c
+
+void aom_smooth_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_16x8 aom_smooth_h_predictor_16x8_c
+
+void aom_smooth_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
+
+void aom_smooth_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_32x16 aom_smooth_h_predictor_32x16_c
+
+void aom_smooth_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_32x32 aom_smooth_h_predictor_32x32_c
+
+void aom_smooth_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_4x4 aom_smooth_h_predictor_4x4_c
+
+void aom_smooth_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_4x8 aom_smooth_h_predictor_4x8_c
+
+void aom_smooth_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_8x16 aom_smooth_h_predictor_8x16_c
+
+void aom_smooth_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_8x4 aom_smooth_h_predictor_8x4_c
+
+void aom_smooth_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_c
+
 void aom_smooth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_smooth_predictor_16x16 aom_smooth_predictor_16x16_c
 
+void aom_smooth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_16x32 aom_smooth_predictor_16x32_c
+
+void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_16x8 aom_smooth_predictor_16x8_c
+
 void aom_smooth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
 
+void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_32x16 aom_smooth_predictor_32x16_c
+
 void aom_smooth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_smooth_predictor_32x32 aom_smooth_predictor_32x32_c
 
 void aom_smooth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_smooth_predictor_4x4 aom_smooth_predictor_4x4_c
 
+void aom_smooth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_4x8 aom_smooth_predictor_4x8_c
+
+void aom_smooth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_8x16 aom_smooth_predictor_8x16_c
+
+void aom_smooth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_8x4 aom_smooth_predictor_8x4_c
+
 void aom_smooth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_smooth_predictor_8x8 aom_smooth_predictor_8x8_c
 
+void aom_smooth_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_c
+
+void aom_smooth_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_16x32 aom_smooth_v_predictor_16x32_c
+
+void aom_smooth_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_16x8 aom_smooth_v_predictor_16x8_c
+
+void aom_smooth_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
+
+void aom_smooth_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_32x16 aom_smooth_v_predictor_32x16_c
+
+void aom_smooth_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_32x32 aom_smooth_v_predictor_32x32_c
+
+void aom_smooth_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_4x4 aom_smooth_v_predictor_4x4_c
+
+void aom_smooth_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_4x8 aom_smooth_v_predictor_4x8_c
+
+void aom_smooth_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_8x16 aom_smooth_v_predictor_8x16_c
+
+void aom_smooth_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_8x4 aom_smooth_v_predictor_8x4_c
+
+void aom_smooth_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_8x8 aom_smooth_v_predictor_8x8_c
+
 uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_sub_pixel_avg_variance16x16 aom_sub_pixel_avg_variance16x16_c
 
 uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_sub_pixel_avg_variance16x32 aom_sub_pixel_avg_variance16x32_c
 
+uint32_t aom_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance16x4 aom_sub_pixel_avg_variance16x4_c
+
 uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_sub_pixel_avg_variance16x8 aom_sub_pixel_avg_variance16x8_c
 
 uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_sub_pixel_avg_variance32x16 aom_sub_pixel_avg_variance32x16_c
 
 uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_sub_pixel_avg_variance32x32 aom_sub_pixel_avg_variance32x32_c
 
 uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_sub_pixel_avg_variance32x64 aom_sub_pixel_avg_variance32x64_c
 
+uint32_t aom_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance32x8 aom_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance4x16 aom_sub_pixel_avg_variance4x16_c
+
 uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_sub_pixel_avg_variance4x4 aom_sub_pixel_avg_variance4x4_c
 
 uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_sub_pixel_avg_variance4x8 aom_sub_pixel_avg_variance4x8_c
 
 uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_sub_pixel_avg_variance64x32 aom_sub_pixel_avg_variance64x32_c
 
 uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_sub_pixel_avg_variance64x64 aom_sub_pixel_avg_variance64x64_c
 
 uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_sub_pixel_avg_variance8x16 aom_sub_pixel_avg_variance8x16_c
 
+uint32_t aom_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_sub_pixel_avg_variance8x32 aom_sub_pixel_avg_variance8x32_c
+
 uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_sub_pixel_avg_variance8x4 aom_sub_pixel_avg_variance8x4_c
 
 uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_sub_pixel_avg_variance8x8 aom_sub_pixel_avg_variance8x8_c
 
 uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_sub_pixel_variance16x16 aom_sub_pixel_variance16x16_c
 
 uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_sub_pixel_variance16x32 aom_sub_pixel_variance16x32_c
 
+uint32_t aom_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance16x4 aom_sub_pixel_variance16x4_c
+
 uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_sub_pixel_variance16x8 aom_sub_pixel_variance16x8_c
 
 uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_sub_pixel_variance32x16 aom_sub_pixel_variance32x16_c
 
 uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_sub_pixel_variance32x32 aom_sub_pixel_variance32x32_c
 
 uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_sub_pixel_variance32x64 aom_sub_pixel_variance32x64_c
 
+uint32_t aom_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance32x8 aom_sub_pixel_variance32x8_c
+
+uint32_t aom_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance4x16 aom_sub_pixel_variance4x16_c
+
 uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_sub_pixel_variance4x4 aom_sub_pixel_variance4x4_c
 
 uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_sub_pixel_variance4x8 aom_sub_pixel_variance4x8_c
 
 uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_sub_pixel_variance64x32 aom_sub_pixel_variance64x32_c
 
 uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_sub_pixel_variance64x64 aom_sub_pixel_variance64x64_c
 
 uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_sub_pixel_variance8x16 aom_sub_pixel_variance8x16_c
 
+uint32_t aom_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_sub_pixel_variance8x32 aom_sub_pixel_variance8x32_c
+
 uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_sub_pixel_variance8x4 aom_sub_pixel_variance8x4_c
 
 uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_sub_pixel_variance8x8 aom_sub_pixel_variance8x8_c
 
 void aom_subtract_block_c(int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride);
 #define aom_subtract_block aom_subtract_block_c
 
 uint64_t aom_sum_squares_2d_i16_c(const int16_t *src, int stride, int width, int height);
 #define aom_sum_squares_2d_i16 aom_sum_squares_2d_i16_c
 
 uint64_t aom_sum_squares_i16_c(const int16_t *src, uint32_t N);
 #define aom_sum_squares_i16 aom_sum_squares_i16_c
 
-void aom_upsampled_pred_c(uint8_t *comp_pred, int width, int height, const uint8_t *ref, int ref_stride);
+void aom_upsampled_pred_c(uint8_t *comp_pred, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref, int ref_stride);
 #define aom_upsampled_pred aom_upsampled_pred_c
 
 void aom_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_v_predictor_16x16 aom_v_predictor_16x16_c
 
+void aom_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x32 aom_v_predictor_16x32_c
+
+void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x8 aom_v_predictor_16x8_c
+
 void aom_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_v_predictor_2x2 aom_v_predictor_2x2_c
 
+void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_32x16 aom_v_predictor_32x16_c
+
 void aom_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_v_predictor_32x32 aom_v_predictor_32x32_c
 
 void aom_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_v_predictor_4x4 aom_v_predictor_4x4_c
 
+void aom_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x8 aom_v_predictor_4x8_c
+
+void aom_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x16 aom_v_predictor_8x16_c
+
+void aom_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x4 aom_v_predictor_8x4_c
+
 void aom_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_v_predictor_8x8 aom_v_predictor_8x8_c
 
 unsigned int aom_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_variance16x16 aom_variance16x16_c
 
 unsigned int aom_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_variance16x32 aom_variance16x32_c
 
+unsigned int aom_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance16x4 aom_variance16x4_c
+
 unsigned int aom_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_variance16x8 aom_variance16x8_c
 
 unsigned int aom_variance2x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_variance2x2 aom_variance2x2_c
 
 unsigned int aom_variance2x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_variance2x4 aom_variance2x4_c
@@ -2307,16 +3321,22 @@ unsigned int aom_variance32x16_c(const u
 #define aom_variance32x16 aom_variance32x16_c
 
 unsigned int aom_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_variance32x32 aom_variance32x32_c
 
 unsigned int aom_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_variance32x64 aom_variance32x64_c
 
+unsigned int aom_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance32x8 aom_variance32x8_c
+
+unsigned int aom_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance4x16 aom_variance4x16_c
+
 unsigned int aom_variance4x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_variance4x2 aom_variance4x2_c
 
 unsigned int aom_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_variance4x4 aom_variance4x4_c
 
 unsigned int aom_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_variance4x8 aom_variance4x8_c
@@ -2325,16 +3345,19 @@ unsigned int aom_variance64x32_c(const u
 #define aom_variance64x32 aom_variance64x32_c
 
 unsigned int aom_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_variance64x64 aom_variance64x64_c
 
 unsigned int aom_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_variance8x16 aom_variance8x16_c
 
+unsigned int aom_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define aom_variance8x32 aom_variance8x32_c
+
 unsigned int aom_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_variance8x4 aom_variance8x4_c
 
 unsigned int aom_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_variance8x8 aom_variance8x8_c
 
 uint32_t aom_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, uint32_t *sse);
 #define aom_variance_halfpixvar16x16_h aom_variance_halfpixvar16x16_h_c
--- a/media/libaom/config/generic/av1_rtcd.h
+++ b/media/libaom/config/generic/av1_rtcd.h
@@ -7,28 +7,30 @@
 #define RTCD_EXTERN extern
 #endif
 
 /*
  * AV1
  */
 
 #include "aom/aom_integer.h"
+#include "aom_dsp/txfm_common.h"
 #include "av1/common/common.h"
 #include "av1/common/enums.h"
 #include "av1/common/quant_common.h"
 #include "av1/common/filter.h"
 #include "av1/common/convolve.h"
 #include "av1/common/av1_txfm.h"
 #include "av1/common/odintrin.h"
 
 struct macroblockd;
 
 /* Encoder forward decls */
 struct macroblock;
+struct txfm_param;
 struct aom_variance_vtable;
 struct search_site_config;
 struct mv;
 union int_mv;
 struct yv12_buffer_config;
 typedef uint16_t od_dering_in;
 
 #ifdef __cplusplus
@@ -57,79 +59,97 @@ void av1_convolve_vert_c(const uint8_t *
 #define av1_convolve_vert av1_convolve_vert_c
 
 int av1_diamond_search_sad_c(struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define av1_diamond_search_sad av1_diamond_search_sad_c
 
 void av1_fdct8x8_quant_c(const int16_t *input, int stride, tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 #define av1_fdct8x8_quant av1_fdct8x8_quant_c
 
-void av1_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht16x16 av1_fht16x16_c
 
-void av1_fht16x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht16x32_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht16x32 av1_fht16x32_c
 
-void av1_fht16x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht16x4_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht16x4 av1_fht16x4_c
 
-void av1_fht16x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht16x8_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht16x8 av1_fht16x8_c
 
-void av1_fht32x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht32x16_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht32x16 av1_fht32x16_c
 
-void av1_fht32x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht32x32_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht32x32 av1_fht32x32_c
 
-void av1_fht32x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht32x8_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht32x8 av1_fht32x8_c
 
-void av1_fht4x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht4x16_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht4x16 av1_fht4x16_c
 
-void av1_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht4x4 av1_fht4x4_c
 
-void av1_fht4x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht4x8_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht4x8 av1_fht4x8_c
 
-void av1_fht8x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht8x16_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht8x16 av1_fht8x16_c
 
-void av1_fht8x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht8x32_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht8x32 av1_fht8x32_c
 
-void av1_fht8x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht8x4_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht8x4 av1_fht8x4_c
 
-void av1_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
+void av1_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht8x8 av1_fht8x8_c
 
 int av1_full_range_search_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define av1_full_range_search av1_full_range_search_c
 
 int av1_full_search_sad_c(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
 #define av1_full_search_sad av1_full_search_sad_c
 
 void av1_fwd_idtx_c(const int16_t *src_diff, tran_low_t *coeff, int stride, int bs, int tx_type);
 #define av1_fwd_idtx av1_fwd_idtx_c
 
 void av1_fwd_txfm2d_16x16_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 #define av1_fwd_txfm2d_16x16 av1_fwd_txfm2d_16x16_c
 
+void av1_fwd_txfm2d_16x32_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+#define av1_fwd_txfm2d_16x32 av1_fwd_txfm2d_16x32_c
+
+void av1_fwd_txfm2d_16x8_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+#define av1_fwd_txfm2d_16x8 av1_fwd_txfm2d_16x8_c
+
+void av1_fwd_txfm2d_32x16_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+#define av1_fwd_txfm2d_32x16 av1_fwd_txfm2d_32x16_c
+
 void av1_fwd_txfm2d_32x32_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 #define av1_fwd_txfm2d_32x32 av1_fwd_txfm2d_32x32_c
 
 void av1_fwd_txfm2d_4x4_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 #define av1_fwd_txfm2d_4x4 av1_fwd_txfm2d_4x4_c
 
+void av1_fwd_txfm2d_4x8_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+#define av1_fwd_txfm2d_4x8 av1_fwd_txfm2d_4x8_c
+
 void av1_fwd_txfm2d_64x64_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 #define av1_fwd_txfm2d_64x64 av1_fwd_txfm2d_64x64_c
 
+void av1_fwd_txfm2d_8x16_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+#define av1_fwd_txfm2d_8x16 av1_fwd_txfm2d_8x16_c
+
+void av1_fwd_txfm2d_8x4_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+#define av1_fwd_txfm2d_8x4 av1_fwd_txfm2d_8x4_c
+
 void av1_fwd_txfm2d_8x8_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
 #define av1_fwd_txfm2d_8x8 av1_fwd_txfm2d_8x8_c
 
 void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
 #define av1_fwht4x4 av1_fwht4x4_c
 
 int64_t av1_highbd_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz, int bd);
 #define av1_highbd_block_error av1_highbd_block_error_c
@@ -162,166 +182,142 @@ void av1_highbd_convolve_horiz_c(const u
 #define av1_highbd_convolve_horiz av1_highbd_convolve_horiz_c
 
 void av1_highbd_convolve_init_c(void);
 #define av1_highbd_convolve_init av1_highbd_convolve_init_c
 
 void av1_highbd_convolve_vert_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd);
 #define av1_highbd_convolve_vert av1_highbd_convolve_vert_c
 
-void av1_highbd_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht16x16 av1_highbd_fht16x16_c
-
-void av1_highbd_fht16x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht16x32 av1_highbd_fht16x32_c
-
-void av1_highbd_fht16x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht16x4 av1_highbd_fht16x4_c
-
-void av1_highbd_fht16x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht16x8 av1_highbd_fht16x8_c
-
-void av1_highbd_fht32x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht32x16 av1_highbd_fht32x16_c
-
-void av1_highbd_fht32x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht32x32 av1_highbd_fht32x32_c
-
-void av1_highbd_fht32x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht32x8 av1_highbd_fht32x8_c
-
-void av1_highbd_fht4x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht4x16 av1_highbd_fht4x16_c
-
-void av1_highbd_fht4x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht4x4 av1_highbd_fht4x4_c
-
-void av1_highbd_fht4x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht4x8 av1_highbd_fht4x8_c
-
-void av1_highbd_fht8x16_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht8x16 av1_highbd_fht8x16_c
-
-void av1_highbd_fht8x32_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht8x32 av1_highbd_fht8x32_c
-
-void av1_highbd_fht8x4_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht8x4 av1_highbd_fht8x4_c
-
-void av1_highbd_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, int tx_type);
-#define av1_highbd_fht8x8 av1_highbd_fht8x8_c
-
 void av1_highbd_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
 #define av1_highbd_fwht4x4 av1_highbd_fwht4x4_c
 
-void av1_highbd_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type, int bd);
+void av1_highbd_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param);
 #define av1_highbd_iht16x16_256_add av1_highbd_iht16x16_256_add_c
 
-void av1_highbd_iht16x32_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht16x32_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht16x32_512_add av1_highbd_iht16x32_512_add_c
 
-void av1_highbd_iht16x4_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht16x4_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht16x4_64_add av1_highbd_iht16x4_64_add_c
 
-void av1_highbd_iht16x8_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht16x8_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht16x8_128_add av1_highbd_iht16x8_128_add_c
 
-void av1_highbd_iht32x16_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht32x16_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht32x16_512_add av1_highbd_iht32x16_512_add_c
 
-void av1_highbd_iht32x8_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht32x8_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht32x8_256_add av1_highbd_iht32x8_256_add_c
 
-void av1_highbd_iht4x16_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht4x16_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht4x16_64_add av1_highbd_iht4x16_64_add_c
 
-void av1_highbd_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht4x4_16_add av1_highbd_iht4x4_16_add_c
 
-void av1_highbd_iht4x8_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht4x8_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht4x8_32_add av1_highbd_iht4x8_32_add_c
 
-void av1_highbd_iht8x16_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht8x16_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht8x16_128_add av1_highbd_iht8x16_128_add_c
 
-void av1_highbd_iht8x32_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht8x32_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht8x32_256_add av1_highbd_iht8x32_256_add_c
 
-void av1_highbd_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht8x4_32_add av1_highbd_iht8x4_32_add_c
 
-void av1_highbd_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type, int bd);
+void av1_highbd_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht8x8_64_add av1_highbd_iht8x8_64_add_c
 
 void av1_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
 #define av1_highbd_quantize_b av1_highbd_quantize_b_c
 
 void av1_highbd_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
 #define av1_highbd_quantize_fp av1_highbd_quantize_fp_c
 
 void av1_highbd_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 #define av1_highbd_temporal_filter_apply av1_highbd_temporal_filter_apply_c
 
-void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, int comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
 #define av1_highbd_warp_affine av1_highbd_warp_affine_c
 
-void av1_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+void av1_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param);
 #define av1_iht16x16_256_add av1_iht16x16_256_add_c
 
-void av1_iht16x32_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht16x32_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht16x32_512_add av1_iht16x32_512_add_c
 
-void av1_iht16x4_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht16x4_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht16x4_64_add av1_iht16x4_64_add_c
 
-void av1_iht16x8_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht16x8_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht16x8_128_add av1_iht16x8_128_add_c
 
-void av1_iht32x16_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht32x16_512_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht32x16_512_add av1_iht32x16_512_add_c
 
-void av1_iht32x32_1024_add_c(const tran_low_t *input, uint8_t *output, int pitch, int tx_type);
+void av1_iht32x32_1024_add_c(const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param);
 #define av1_iht32x32_1024_add av1_iht32x32_1024_add_c
 
-void av1_iht32x8_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht32x8_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht32x8_256_add av1_iht32x8_256_add_c
 
-void av1_iht4x16_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht4x16_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht4x16_64_add av1_iht4x16_64_add_c
 
-void av1_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht4x4_16_add av1_iht4x4_16_add_c
 
-void av1_iht4x8_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht4x8_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht4x8_32_add av1_iht4x8_32_add_c
 
-void av1_iht8x16_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht8x16_128_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht8x16_128_add av1_iht8x16_128_add_c
 
-void av1_iht8x32_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht8x32_256_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht8x32_256_add av1_iht8x32_256_add_c
 
-void av1_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht8x4_32_add av1_iht8x4_32_add_c
 
-void av1_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void av1_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht8x8_64_add av1_iht8x8_64_add_c
 
 void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 #define av1_inv_txfm2d_add_16x16 av1_inv_txfm2d_add_16x16_c
 
+void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+#define av1_inv_txfm2d_add_16x32 av1_inv_txfm2d_add_16x32_c
+
+void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+#define av1_inv_txfm2d_add_16x8 av1_inv_txfm2d_add_16x8_c
+
+void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+#define av1_inv_txfm2d_add_32x16 av1_inv_txfm2d_add_32x16_c
+
 void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 #define av1_inv_txfm2d_add_32x32 av1_inv_txfm2d_add_32x32_c
 
 void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 #define av1_inv_txfm2d_add_4x4 av1_inv_txfm2d_add_4x4_c
 
+void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+#define av1_inv_txfm2d_add_4x8 av1_inv_txfm2d_add_4x8_c
+
 void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 #define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c
 
+void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+#define av1_inv_txfm2d_add_8x16 av1_inv_txfm2d_add_8x16_c
+
+void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+#define av1_inv_txfm2d_add_8x4 av1_inv_txfm2d_add_8x4_c
+
 void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
 #define av1_inv_txfm2d_add_8x8 av1_inv_txfm2d_add_8x8_c
 
 void av1_lowbd_convolve_init_c(void);
 #define av1_lowbd_convolve_init av1_lowbd_convolve_init_c
 
 void av1_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan, int log_scale);
 #define av1_quantize_b av1_quantize_b_c
@@ -330,17 +326,17 @@ void av1_quantize_fp_c(const tran_low_t 
 #define av1_quantize_fp av1_quantize_fp_c
 
 void av1_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 #define av1_quantize_fp_32x32 av1_quantize_fp_32x32_c
 
 void av1_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 #define av1_temporal_filter_apply av1_temporal_filter_apply_c
 
-void av1_warp_affine_c(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int comp_avg, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+void av1_warp_affine_c(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
 #define av1_warp_affine av1_warp_affine_c
 
 void av1_wedge_compute_delta_squares_c(int16_t *d, const int16_t *a, const int16_t *b, int N);
 #define av1_wedge_compute_delta_squares av1_wedge_compute_delta_squares_c
 
 int av1_wedge_sign_from_residuals_c(const int16_t *ds, const uint8_t *m, int N, int64_t limit);
 #define av1_wedge_sign_from_residuals av1_wedge_sign_from_residuals_c
 
--- a/media/libaom/config/linux/arm/aom_config.asm
+++ b/media/libaom/config/linux/arm/aom_config.asm
@@ -63,16 +63,17 @@
 .equ CONFIG_INSPECTION ,  0
 .equ CONFIG_DECODE_PERF_TESTS ,  0
 .equ CONFIG_ENCODE_PERF_TESTS ,  0
 .equ CONFIG_COEFFICIENT_RANGE_CHECKING ,  0
 .equ CONFIG_LOWBITDEPTH ,  1
 .equ CONFIG_HIGHBITDEPTH ,  1
 .equ CONFIG_EXPERIMENTAL ,  0
 .equ CONFIG_SIZE_LIMIT ,  1
+.equ CONFIG_COLORSPACE_HEADERS ,  0
 .equ CONFIG_FP_MB_STATS ,  0
 .equ CONFIG_CDEF ,  1
 .equ CONFIG_VAR_TX ,  1
 .equ CONFIG_RECT_TX ,  1
 .equ CONFIG_RECT_TX_EXT ,  0
 .equ CONFIG_TPL_MV ,  0
 .equ CONFIG_DUAL_FILTER ,  1
 .equ CONFIG_CONVOLVE_ROUND ,  0
@@ -85,66 +86,82 @@
 .equ CONFIG_FILTER_INTRA ,  0
 .equ CONFIG_INTRA_EDGE ,  0
 .equ CONFIG_INTRABC ,  0
 .equ CONFIG_EXT_INTER ,  1
 .equ CONFIG_INTERINTRA ,  1
 .equ CONFIG_WEDGE ,  1
 .equ CONFIG_COMPOUND_SEGMENT ,  1
 .equ CONFIG_EXT_REFS ,  1
+.equ CONFIG_ALTREF2 ,  0
+.equ CONFIG_SPEED_REFS ,  0
+.equ CONFIG_GF_GROUPS ,  0
+.equ CONFIG_FLEX_REFS ,  0
 .equ CONFIG_GLOBAL_MOTION ,  1
 .equ CONFIG_NEW_QUANT ,  0
 .equ CONFIG_SUPERTX ,  0
 .equ CONFIG_ANS ,  0
 .equ CONFIG_LOOP_RESTORATION ,  0
 .equ CONFIG_EXT_PARTITION ,  0
 .equ CONFIG_EXT_PARTITION_TYPES ,  0
 .equ CONFIG_UNPOISON_PARTITION_CTX ,  0
 .equ CONFIG_EXT_TILE ,  0
 .equ CONFIG_MOTION_VAR ,  1
 .equ CONFIG_NCOBMC ,  0
 .equ CONFIG_WARPED_MOTION ,  1
 .equ CONFIG_Q_ADAPT_PROBS ,  0
 .equ CONFIG_BITSTREAM_DEBUG ,  0
+.equ CONFIG_INTER_STATS_ONLY ,  0
 .equ CONFIG_ALT_INTRA ,  1
 .equ CONFIG_PALETTE ,  1
 .equ CONFIG_PALETTE_DELTA_ENCODING ,  0
 .equ CONFIG_RAWBITS ,  0
 .equ CONFIG_EC_SMALLMUL ,  1
 .equ CONFIG_PVQ ,  0
 .equ CONFIG_CFL ,  0
 .equ CONFIG_XIPHRC ,  0
+.equ CONFIG_DCT_ONLY ,  0
+.equ CONFIG_DAALA_DCT4 ,  0
+.equ CONFIG_DAALA_DCT8 ,  0
 .equ CONFIG_CB4X4 ,  1
 .equ CONFIG_CHROMA_2X2 ,  0
 .equ CONFIG_CHROMA_SUB8X8 ,  1
 .equ CONFIG_FRAME_SIZE ,  0
 .equ CONFIG_DELTA_Q ,  1
 .equ CONFIG_EXT_DELTA_Q ,  1
 .equ CONFIG_ADAPT_SCAN ,  0
 .equ CONFIG_FILTER_7BIT ,  1
-.equ CONFIG_PARALLEL_DEBLOCKING ,  0
-.equ CONFIG_PARALLEL_DEBLOCKING_15TAP ,  0
+.equ CONFIG_PARALLEL_DEBLOCKING ,  1
 .equ CONFIG_LOOPFILTERING_ACROSS_TILES ,  1
-.equ CONFIG_TILE_GROUPS ,  1
-.equ CONFIG_EC_ADAPT ,  1
 .equ CONFIG_TEMPMV_SIGNALING ,  1
 .equ CONFIG_RD_DEBUG ,  0
 .equ CONFIG_REFERENCE_BUFFER ,  1
 .equ CONFIG_COEF_INTERLEAVE ,  0
 .equ CONFIG_ENTROPY_STATS ,  0
 .equ CONFIG_MASKED_TX ,  0
 .equ CONFIG_DEPENDENT_HORZTILES ,  0
+.equ CONFIG_DIST_8X8 ,  0
 .equ CONFIG_DAALA_DIST ,  0
 .equ CONFIG_TRIPRED ,  0
 .equ CONFIG_PALETTE_THROUGHPUT ,  1
 .equ CONFIG_REF_ADAPT ,  0
 .equ CONFIG_LV_MAP ,  0
 .equ CONFIG_TXK_SEL ,  0
 .equ CONFIG_MV_COMPRESS ,  1
 .equ CONFIG_FRAME_SUPERRES ,  0
 .equ CONFIG_NEW_MULTISYMBOL ,  0
 .equ CONFIG_COMPOUND_SINGLEREF ,  0
 .equ CONFIG_AOM_QM ,  0
 .equ CONFIG_ONE_SIDED_COMPOUND ,  1
-.equ CONFIG_SMOOTH_HV ,  0
+.equ CONFIG_EXT_COMP_REFS ,  0
+.equ CONFIG_SMOOTH_HV ,  1
 .equ CONFIG_VAR_REFS ,  0
+.equ CONFIG_RECT_INTRA_PRED ,  1
+.equ CONFIG_LGT ,  0
+.equ CONFIG_SBL_SYMBOL ,  0
+.equ CONFIG_NCOBMC_ADAPT_WEIGHT ,  0
+.equ CONFIG_BGSPRITE ,  0
+.equ CONFIG_VAR_TX_NO_TX_MODE ,  0
+.equ CONFIG_MRC_TX ,  0
+.equ CONFIG_LPF_DIRECT ,  0
+.equ CONFIG_UV_LVL ,  0
 .equ CONFIG_ANALYZER ,  0
 	.section	.note.GNU-stack,"",%progbits
--- a/media/libaom/config/linux/arm/aom_config.h
+++ b/media/libaom/config/linux/arm/aom_config.h
@@ -4,17 +4,17 @@
 /* the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License */
 /* was not distributed with this source code in the LICENSE file, you can */
 /* obtain it at www.aomedia.org/license/software. If the Alliance for Open */
 /* Media Patent License 1.0 was not distributed with this source code in the */
 /* PATENTS file, you can obtain it at www.aomedia.org/license/patent. */
 /* This file automatically generated by configure. Do not edit! */
 #ifndef AOM_CONFIG_H
 #define AOM_CONFIG_H
-#define RESTRICT
+#define RESTRICT    
 #define INLINE      inline
 #define ARCH_ARM 1
 #define ARCH_MIPS 0
 #define ARCH_X86 0
 #define ARCH_X86_64 0
 #define HAVE_NEON 1
 #define HAVE_NEON_ASM 1
 #define HAVE_MIPS32 0
@@ -73,16 +73,17 @@
 #define CONFIG_INSPECTION 0
 #define CONFIG_DECODE_PERF_TESTS 0
 #define CONFIG_ENCODE_PERF_TESTS 0
 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0
 #define CONFIG_LOWBITDEPTH 1
 #define CONFIG_HIGHBITDEPTH 1
 #define CONFIG_EXPERIMENTAL 0
 #define CONFIG_SIZE_LIMIT 1
+#define CONFIG_COLORSPACE_HEADERS 0
 #define CONFIG_FP_MB_STATS 0
 #define CONFIG_CDEF 1
 #define CONFIG_VAR_TX 1
 #define CONFIG_RECT_TX 1
 #define CONFIG_RECT_TX_EXT 0
 #define CONFIG_TPL_MV 0
 #define CONFIG_DUAL_FILTER 1
 #define CONFIG_CONVOLVE_ROUND 0
@@ -95,68 +96,84 @@
 #define CONFIG_FILTER_INTRA 0
 #define CONFIG_INTRA_EDGE 0
 #define CONFIG_INTRABC 0
 #define CONFIG_EXT_INTER 1
 #define CONFIG_INTERINTRA 1
 #define CONFIG_WEDGE 1
 #define CONFIG_COMPOUND_SEGMENT 1
 #define CONFIG_EXT_REFS 1
+#define CONFIG_ALTREF2 0
+#define CONFIG_SPEED_REFS 0
+#define CONFIG_GF_GROUPS 0
+#define CONFIG_FLEX_REFS 0
 #define CONFIG_GLOBAL_MOTION 1
 #define CONFIG_NEW_QUANT 0
 #define CONFIG_SUPERTX 0
 #define CONFIG_ANS 0
 #define CONFIG_LOOP_RESTORATION 0
 #define CONFIG_EXT_PARTITION 0
 #define CONFIG_EXT_PARTITION_TYPES 0
 #define CONFIG_UNPOISON_PARTITION_CTX 0
 #define CONFIG_EXT_TILE 0
 #define CONFIG_MOTION_VAR 1
 #define CONFIG_NCOBMC 0
 #define CONFIG_WARPED_MOTION 1
 #define CONFIG_Q_ADAPT_PROBS 0
 #define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_INTER_STATS_ONLY 0
 #define CONFIG_ALT_INTRA 1
 #define CONFIG_PALETTE 1
 #define CONFIG_PALETTE_DELTA_ENCODING 0
 #define CONFIG_RAWBITS 0
 #define CONFIG_EC_SMALLMUL 1
 #define CONFIG_PVQ 0
 #define CONFIG_CFL 0
 #define CONFIG_XIPHRC 0
+#define CONFIG_DCT_ONLY 0
+#define CONFIG_DAALA_DCT4 0
+#define CONFIG_DAALA_DCT8 0
 #define CONFIG_CB4X4 1
 #define CONFIG_CHROMA_2X2 0
 #define CONFIG_CHROMA_SUB8X8 1
 #define CONFIG_FRAME_SIZE 0
 #define CONFIG_DELTA_Q 1
 #define CONFIG_EXT_DELTA_Q 1
 #define CONFIG_ADAPT_SCAN 0
 #define CONFIG_FILTER_7BIT 1
-#define CONFIG_PARALLEL_DEBLOCKING 0
-#define CONFIG_PARALLEL_DEBLOCKING_15TAP 0
+#define CONFIG_PARALLEL_DEBLOCKING 1
 #define CONFIG_LOOPFILTERING_ACROSS_TILES 1
-#define CONFIG_TILE_GROUPS 1
-#define CONFIG_EC_ADAPT 1
 #define CONFIG_TEMPMV_SIGNALING 1
 #define CONFIG_RD_DEBUG 0
 #define CONFIG_REFERENCE_BUFFER 1
 #define CONFIG_COEF_INTERLEAVE 0
 #define CONFIG_ENTROPY_STATS 0
 #define CONFIG_MASKED_TX 0
 #define CONFIG_DEPENDENT_HORZTILES 0
+#define CONFIG_DIST_8X8 0
 #define CONFIG_DAALA_DIST 0
 #define CONFIG_TRIPRED 0
 #define CONFIG_PALETTE_THROUGHPUT 1
 #define CONFIG_REF_ADAPT 0
 #define CONFIG_LV_MAP 0
 #define CONFIG_TXK_SEL 0
 #define CONFIG_MV_COMPRESS 1
 #define CONFIG_FRAME_SUPERRES 0
 #define CONFIG_NEW_MULTISYMBOL 0
 #define CONFIG_COMPOUND_SINGLEREF 0
 #define CONFIG_AOM_QM 0
 #define CONFIG_ONE_SIDED_COMPOUND 1
-#define CONFIG_SMOOTH_HV 0
+#define CONFIG_EXT_COMP_REFS 0
+#define CONFIG_SMOOTH_HV 1
 #define CONFIG_VAR_REFS 0
+#define CONFIG_RECT_INTRA_PRED 1
+#define CONFIG_LGT 0
+#define CONFIG_SBL_SYMBOL 0
+#define CONFIG_NCOBMC_ADAPT_WEIGHT 0
+#define CONFIG_BGSPRITE 0
+#define CONFIG_VAR_TX_NO_TX_MODE 0
+#define CONFIG_MRC_TX 0
+#define CONFIG_LPF_DIRECT 0
+#define CONFIG_UV_LVL 0
 #define CONFIG_ANALYZER 0
 #define DECODE_WIDTH_LIMIT 8192
 #define DECODE_HEIGHT_LIMIT 4608
 #endif /* AOM_CONFIG_H */
--- a/media/libaom/config/linux/arm/aom_dsp_rtcd.h
+++ b/media/libaom/config/linux/arm/aom_dsp_rtcd.h
@@ -27,251 +27,440 @@ void aom_blend_a64_mask_c(uint8_t *dst, 
 #define aom_blend_a64_mask aom_blend_a64_mask_c
 
 void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w);
 #define aom_blend_a64_vmask aom_blend_a64_vmask_c
 
 void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
 #define aom_comp_avg_pred aom_comp_avg_pred_c
 
-void aom_comp_avg_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride);
+void aom_comp_avg_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref, int ref_stride);
 #define aom_comp_avg_upsampled_pred aom_comp_avg_upsampled_pred_c
 
 void aom_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 #define aom_comp_mask_pred aom_comp_mask_pred_c
 
-void aom_comp_mask_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_comp_mask_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 #define aom_comp_mask_upsampled_pred aom_comp_mask_upsampled_pred_c
 
 void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void aom_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_avg)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void aom_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_avg_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_avg_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
+void aom_convolve8_avg_horiz_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_avg_horiz_scale aom_convolve8_avg_horiz_scale_c
+
+void aom_convolve8_avg_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_avg_scale aom_convolve8_avg_scale_c
+
 void aom_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_avg_vert_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_avg_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
+void aom_convolve8_avg_vert_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_avg_vert_scale aom_convolve8_avg_vert_scale_c
+
 void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
+void aom_convolve8_horiz_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_horiz_scale aom_convolve8_horiz_scale_c
+
+void aom_convolve8_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_scale aom_convolve8_scale_c
+
 void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
+void aom_convolve8_vert_scale_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int subpel_x, int x_step_q4, const int16_t *filter_y, int subpel_y, int y_step_q4, int w, int h);
+#define aom_convolve8_vert_scale aom_convolve8_vert_scale_c
+
 void aom_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve_avg_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve_avg)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void aom_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve_copy_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void aom_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d117_predictor_16x16 aom_d117_predictor_16x16_c
 
+void aom_d117_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_16x32 aom_d117_predictor_16x32_c
+
+void aom_d117_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_16x8 aom_d117_predictor_16x8_c
+
 void aom_d117_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d117_predictor_2x2 aom_d117_predictor_2x2_c
 
+void aom_d117_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_32x16 aom_d117_predictor_32x16_c
+
 void aom_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d117_predictor_32x32 aom_d117_predictor_32x32_c
 
 void aom_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d117_predictor_4x4 aom_d117_predictor_4x4_c
 
+void aom_d117_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_4x8 aom_d117_predictor_4x8_c
+
+void aom_d117_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_8x16 aom_d117_predictor_8x16_c
+
+void aom_d117_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d117_predictor_8x4 aom_d117_predictor_8x4_c
+
 void aom_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d117_predictor_8x8 aom_d117_predictor_8x8_c
 
 void aom_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d135_predictor_16x16 aom_d135_predictor_16x16_c
 
+void aom_d135_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_16x32 aom_d135_predictor_16x32_c
+
+void aom_d135_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_16x8 aom_d135_predictor_16x8_c
+
 void aom_d135_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d135_predictor_2x2 aom_d135_predictor_2x2_c
 
+void aom_d135_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_32x16 aom_d135_predictor_32x16_c
+
 void aom_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d135_predictor_32x32 aom_d135_predictor_32x32_c
 
 void aom_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_d135_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
+void aom_d135_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_4x8 aom_d135_predictor_4x8_c
+
+void aom_d135_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_8x16 aom_d135_predictor_8x16_c
+
+void aom_d135_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d135_predictor_8x4 aom_d135_predictor_8x4_c
+
 void aom_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d135_predictor_8x8 aom_d135_predictor_8x8_c
 
 void aom_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d153_predictor_16x16 aom_d153_predictor_16x16_c
 
+void aom_d153_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_16x32 aom_d153_predictor_16x32_c
+
+void aom_d153_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_16x8 aom_d153_predictor_16x8_c
+
 void aom_d153_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d153_predictor_2x2 aom_d153_predictor_2x2_c
 
+void aom_d153_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_32x16 aom_d153_predictor_32x16_c
+
 void aom_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d153_predictor_32x32 aom_d153_predictor_32x32_c
 
 void aom_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d153_predictor_4x4 aom_d153_predictor_4x4_c
 
+void aom_d153_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_4x8 aom_d153_predictor_4x8_c
+
+void aom_d153_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_8x16 aom_d153_predictor_8x16_c
+
+void aom_d153_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d153_predictor_8x4 aom_d153_predictor_8x4_c
+
 void aom_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d153_predictor_8x8 aom_d153_predictor_8x8_c
 
 void aom_d207e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d207e_predictor_16x16 aom_d207e_predictor_16x16_c
 
+void aom_d207e_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_16x32 aom_d207e_predictor_16x32_c
+
+void aom_d207e_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_16x8 aom_d207e_predictor_16x8_c
+
 void aom_d207e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d207e_predictor_2x2 aom_d207e_predictor_2x2_c
 
+void aom_d207e_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_32x16 aom_d207e_predictor_32x16_c
+
 void aom_d207e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d207e_predictor_32x32 aom_d207e_predictor_32x32_c
 
 void aom_d207e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d207e_predictor_4x4 aom_d207e_predictor_4x4_c
 
+void aom_d207e_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_4x8 aom_d207e_predictor_4x8_c
+
+void aom_d207e_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_8x16 aom_d207e_predictor_8x16_c
+
+void aom_d207e_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d207e_predictor_8x4 aom_d207e_predictor_8x4_c
+
 void aom_d207e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d207e_predictor_8x8 aom_d207e_predictor_8x8_c
 
 void aom_d45e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d45e_predictor_16x16 aom_d45e_predictor_16x16_c
 
+void aom_d45e_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_16x32 aom_d45e_predictor_16x32_c
+
+void aom_d45e_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_16x8 aom_d45e_predictor_16x8_c
+
 void aom_d45e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d45e_predictor_2x2 aom_d45e_predictor_2x2_c
 
+void aom_d45e_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_32x16 aom_d45e_predictor_32x16_c
+
 void aom_d45e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d45e_predictor_32x32 aom_d45e_predictor_32x32_c
 
 void aom_d45e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d45e_predictor_4x4 aom_d45e_predictor_4x4_c
 
+void aom_d45e_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_4x8 aom_d45e_predictor_4x8_c
+
+void aom_d45e_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_8x16 aom_d45e_predictor_8x16_c
+
+void aom_d45e_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d45e_predictor_8x4 aom_d45e_predictor_8x4_c
+
 void aom_d45e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d45e_predictor_8x8 aom_d45e_predictor_8x8_c
 
 void aom_d63e_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d63e_predictor_16x16 aom_d63e_predictor_16x16_c
 
+void aom_d63e_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_16x32 aom_d63e_predictor_16x32_c
+
+void aom_d63e_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_16x8 aom_d63e_predictor_16x8_c
+
 void aom_d63e_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d63e_predictor_2x2 aom_d63e_predictor_2x2_c
 
+void aom_d63e_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_32x16 aom_d63e_predictor_32x16_c
+
 void aom_d63e_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d63e_predictor_32x32 aom_d63e_predictor_32x32_c
 
 void aom_d63e_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d63e_predictor_4x4 aom_d63e_predictor_4x4_c
 
+void aom_d63e_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_4x8 aom_d63e_predictor_4x8_c
+
+void aom_d63e_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_8x16 aom_d63e_predictor_8x16_c
+
+void aom_d63e_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_d63e_predictor_8x4 aom_d63e_predictor_8x4_c
+
 void aom_d63e_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_d63e_predictor_8x8 aom_d63e_predictor_8x8_c
 
 void aom_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_128_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_128_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
+void aom_dc_128_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_c
+
+void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c
+
 void aom_dc_128_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
 
+void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_32x16 aom_dc_128_predictor_32x16_c
+
 void aom_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_128_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_128_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_128_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_128_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
+void aom_dc_128_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_c
+
+void aom_dc_128_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_c
+
+void aom_dc_128_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x4 aom_dc_128_predictor_8x4_c
+
 void aom_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_128_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_128_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_left_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_left_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
+void aom_dc_left_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_c
+
+void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c
+
 void aom_dc_left_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
 
+void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_32x16 aom_dc_left_predictor_32x16_c
+
 void aom_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_left_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_left_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_left_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_left_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
+void aom_dc_left_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_c
+
+void aom_dc_left_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_c
+
+void aom_dc_left_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x4 aom_dc_left_predictor_8x4_c
+
 void aom_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_left_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_left_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
+void aom_dc_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x32 aom_dc_predictor_16x32_c
+
+void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c
+
 void aom_dc_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
 
+void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_32x16 aom_dc_predictor_32x16_c
+
 void aom_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
+void aom_dc_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x8 aom_dc_predictor_4x8_c
+
+void aom_dc_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x16 aom_dc_predictor_8x16_c
+
+void aom_dc_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x4 aom_dc_predictor_8x4_c
+
 void aom_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_top_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_top_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
+void aom_dc_top_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_c
+
+void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c
+
 void aom_dc_top_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
 
+void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_32x16 aom_dc_top_predictor_32x16_c
+
 void aom_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_top_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_top_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_top_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_top_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
+void aom_dc_top_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_c
+
+void aom_dc_top_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_c
+
+void aom_dc_top_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x4 aom_dc_top_predictor_8x4_c
+
 void aom_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_top_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_dc_top_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_fdct16x16 aom_fdct16x16_c
 
-void aom_fdct16x16_1_c(const int16_t *input, tran_low_t *output, int stride);
-#define aom_fdct16x16_1 aom_fdct16x16_1_c
-
 void aom_fdct32x32_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_fdct32x32 aom_fdct32x32_c
 
-void aom_fdct32x32_1_c(const int16_t *input, tran_low_t *output, int stride);
-#define aom_fdct32x32_1 aom_fdct32x32_1_c
-
 void aom_fdct32x32_rd_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_fdct32x32_rd aom_fdct32x32_rd_c
 
 void aom_fdct4x4_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_fdct4x4 aom_fdct4x4_c
 
 void aom_fdct4x4_1_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_fdct4x4_1 aom_fdct4x4_1_c
 
 void aom_fdct8x8_c(const int16_t *input, tran_low_t *output, int stride);
 #define aom_fdct8x8 aom_fdct8x8_c
 
-void aom_fdct8x8_1_c(const int16_t *input, tran_low_t *output, int stride);
-#define aom_fdct8x8_1 aom_fdct8x8_1_c
-
 void aom_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
 void aom_get16x16var_neon(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
 RTCD_EXTERN void (*aom_get16x16var)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
 
 unsigned int aom_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride);
 unsigned int aom_get4x4sse_cs_neon(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride);
 RTCD_EXTERN unsigned int (*aom_get4x4sse_cs)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride);
 
@@ -281,27 +470,45 @@ RTCD_EXTERN void (*aom_get8x8var)(const 
 
 unsigned int aom_get_mb_ss_c(const int16_t *);
 #define aom_get_mb_ss aom_get_mb_ss_c
 
 void aom_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
+void aom_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x32 aom_h_predictor_16x32_c
+
+void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x8 aom_h_predictor_16x8_c
+
 void aom_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_h_predictor_2x2 aom_h_predictor_2x2_c
 
+void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x16 aom_h_predictor_32x16_c
+
 void aom_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
+void aom_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x8 aom_h_predictor_4x8_c
+
+void aom_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x16 aom_h_predictor_8x16_c
+
+void aom_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x4 aom_h_predictor_8x4_c
+
 void aom_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 RTCD_EXTERN void (*aom_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_hadamard_16x16_c(const int16_t *src_diff, int src_stride, int16_t *coeff);
 void aom_hadamard_16x16_neon(const int16_t *src_diff, int src_stride, int16_t *coeff);
 RTCD_EXTERN void (*aom_hadamard_16x16)(const int16_t *src_diff, int src_stride, int16_t *coeff);
 
@@ -316,43 +523,55 @@ void aom_highbd_10_get8x8var_c(const uin
 #define aom_highbd_10_get8x8var aom_highbd_10_get8x8var_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance16x16 aom_highbd_10_masked_sub_pixel_variance16x16_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance16x32 aom_highbd_10_masked_sub_pixel_variance16x32_c
 
+unsigned int aom_highbd_10_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance16x4 aom_highbd_10_masked_sub_pixel_variance16x4_c
+
 unsigned int aom_highbd_10_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance16x8 aom_highbd_10_masked_sub_pixel_variance16x8_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance32x16 aom_highbd_10_masked_sub_pixel_variance32x16_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance32x32 aom_highbd_10_masked_sub_pixel_variance32x32_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance32x64 aom_highbd_10_masked_sub_pixel_variance32x64_c
 
+unsigned int aom_highbd_10_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance32x8 aom_highbd_10_masked_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_10_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance4x16 aom_highbd_10_masked_sub_pixel_variance4x16_c
+
 unsigned int aom_highbd_10_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance4x4 aom_highbd_10_masked_sub_pixel_variance4x4_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance4x8 aom_highbd_10_masked_sub_pixel_variance4x8_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance64x32 aom_highbd_10_masked_sub_pixel_variance64x32_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance64x64 aom_highbd_10_masked_sub_pixel_variance64x64_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance8x16 aom_highbd_10_masked_sub_pixel_variance8x16_c
 
+unsigned int aom_highbd_10_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_10_masked_sub_pixel_variance8x32 aom_highbd_10_masked_sub_pixel_variance8x32_c
+
 unsigned int aom_highbd_10_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance8x4 aom_highbd_10_masked_sub_pixel_variance8x4_c
 
 unsigned int aom_highbd_10_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_10_masked_sub_pixel_variance8x8 aom_highbd_10_masked_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define aom_highbd_10_mse16x16 aom_highbd_10_mse16x16_c
@@ -367,172 +586,223 @@ unsigned int aom_highbd_10_mse8x8_c(cons
 #define aom_highbd_10_mse8x8 aom_highbd_10_mse8x8_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance16x16 aom_highbd_10_obmc_sub_pixel_variance16x16_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance16x32 aom_highbd_10_obmc_sub_pixel_variance16x32_c
 
+unsigned int aom_highbd_10_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance16x4 aom_highbd_10_obmc_sub_pixel_variance16x4_c
+
 unsigned int aom_highbd_10_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance16x8 aom_highbd_10_obmc_sub_pixel_variance16x8_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance32x16 aom_highbd_10_obmc_sub_pixel_variance32x16_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance32x32 aom_highbd_10_obmc_sub_pixel_variance32x32_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance32x64 aom_highbd_10_obmc_sub_pixel_variance32x64_c
 
+unsigned int aom_highbd_10_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance32x8 aom_highbd_10_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_10_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance4x16 aom_highbd_10_obmc_sub_pixel_variance4x16_c
+
 unsigned int aom_highbd_10_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance4x4 aom_highbd_10_obmc_sub_pixel_variance4x4_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance4x8 aom_highbd_10_obmc_sub_pixel_variance4x8_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance64x32 aom_highbd_10_obmc_sub_pixel_variance64x32_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance64x64 aom_highbd_10_obmc_sub_pixel_variance64x64_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance8x16 aom_highbd_10_obmc_sub_pixel_variance8x16_c
 
+unsigned int aom_highbd_10_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_sub_pixel_variance8x32 aom_highbd_10_obmc_sub_pixel_variance8x32_c
+
 unsigned int aom_highbd_10_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance8x4 aom_highbd_10_obmc_sub_pixel_variance8x4_c
 
 unsigned int aom_highbd_10_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_sub_pixel_variance8x8 aom_highbd_10_obmc_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_10_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance16x16 aom_highbd_10_obmc_variance16x16_c
 
 unsigned int aom_highbd_10_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance16x32 aom_highbd_10_obmc_variance16x32_c
 
+unsigned int aom_highbd_10_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance16x4 aom_highbd_10_obmc_variance16x4_c
+
 unsigned int aom_highbd_10_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance16x8 aom_highbd_10_obmc_variance16x8_c
 
 unsigned int aom_highbd_10_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance32x16 aom_highbd_10_obmc_variance32x16_c
 
 unsigned int aom_highbd_10_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance32x32 aom_highbd_10_obmc_variance32x32_c
 
 unsigned int aom_highbd_10_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance32x64 aom_highbd_10_obmc_variance32x64_c
 
+unsigned int aom_highbd_10_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance32x8 aom_highbd_10_obmc_variance32x8_c
+
+unsigned int aom_highbd_10_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance4x16 aom_highbd_10_obmc_variance4x16_c
+
 unsigned int aom_highbd_10_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance4x4 aom_highbd_10_obmc_variance4x4_c
 
 unsigned int aom_highbd_10_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance4x8 aom_highbd_10_obmc_variance4x8_c
 
 unsigned int aom_highbd_10_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance64x32 aom_highbd_10_obmc_variance64x32_c
 
 unsigned int aom_highbd_10_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance64x64 aom_highbd_10_obmc_variance64x64_c
 
 unsigned int aom_highbd_10_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance8x16 aom_highbd_10_obmc_variance8x16_c
 
+unsigned int aom_highbd_10_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_10_obmc_variance8x32 aom_highbd_10_obmc_variance8x32_c
+
 unsigned int aom_highbd_10_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance8x4 aom_highbd_10_obmc_variance8x4_c
 
 unsigned int aom_highbd_10_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_10_obmc_variance8x8 aom_highbd_10_obmc_variance8x8_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance16x16 aom_highbd_10_sub_pixel_avg_variance16x16_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance16x32 aom_highbd_10_sub_pixel_avg_variance16x32_c
 
+uint32_t aom_highbd_10_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance16x4 aom_highbd_10_sub_pixel_avg_variance16x4_c
+
 uint32_t aom_highbd_10_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance16x8 aom_highbd_10_sub_pixel_avg_variance16x8_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance32x16 aom_highbd_10_sub_pixel_avg_variance32x16_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance32x32 aom_highbd_10_sub_pixel_avg_variance32x32_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance32x64 aom_highbd_10_sub_pixel_avg_variance32x64_c
 
+uint32_t aom_highbd_10_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance32x8 aom_highbd_10_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance4x16 aom_highbd_10_sub_pixel_avg_variance4x16_c
+
 uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance4x4 aom_highbd_10_sub_pixel_avg_variance4x4_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance4x8 aom_highbd_10_sub_pixel_avg_variance4x8_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance64x32 aom_highbd_10_sub_pixel_avg_variance64x32_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance64x64 aom_highbd_10_sub_pixel_avg_variance64x64_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance8x16 aom_highbd_10_sub_pixel_avg_variance8x16_c
 
+uint32_t aom_highbd_10_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_10_sub_pixel_avg_variance8x32 aom_highbd_10_sub_pixel_avg_variance8x32_c
+
 uint32_t aom_highbd_10_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance8x4 aom_highbd_10_sub_pixel_avg_variance8x4_c
 
 uint32_t aom_highbd_10_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_10_sub_pixel_avg_variance8x8 aom_highbd_10_sub_pixel_avg_variance8x8_c
 
 uint32_t aom_highbd_10_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance16x16 aom_highbd_10_sub_pixel_variance16x16_c
 
 uint32_t aom_highbd_10_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance16x32 aom_highbd_10_sub_pixel_variance16x32_c
 
+uint32_t aom_highbd_10_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance16x4 aom_highbd_10_sub_pixel_variance16x4_c
+
 uint32_t aom_highbd_10_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance16x8 aom_highbd_10_sub_pixel_variance16x8_c
 
 uint32_t aom_highbd_10_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance32x16 aom_highbd_10_sub_pixel_variance32x16_c
 
 uint32_t aom_highbd_10_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance32x32 aom_highbd_10_sub_pixel_variance32x32_c
 
 uint32_t aom_highbd_10_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance32x64 aom_highbd_10_sub_pixel_variance32x64_c
 
+uint32_t aom_highbd_10_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance32x8 aom_highbd_10_sub_pixel_variance32x8_c
+
+uint32_t aom_highbd_10_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance4x16 aom_highbd_10_sub_pixel_variance4x16_c
+
 uint32_t aom_highbd_10_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance4x4 aom_highbd_10_sub_pixel_variance4x4_c
 
 uint32_t aom_highbd_10_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance4x8 aom_highbd_10_sub_pixel_variance4x8_c
 
 uint32_t aom_highbd_10_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance64x32 aom_highbd_10_sub_pixel_variance64x32_c
 
 uint32_t aom_highbd_10_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance64x64 aom_highbd_10_sub_pixel_variance64x64_c
 
 uint32_t aom_highbd_10_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance8x16 aom_highbd_10_sub_pixel_variance8x16_c
 
+uint32_t aom_highbd_10_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_sub_pixel_variance8x32 aom_highbd_10_sub_pixel_variance8x32_c
+
 uint32_t aom_highbd_10_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance8x4 aom_highbd_10_sub_pixel_variance8x4_c
 
 uint32_t aom_highbd_10_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_10_sub_pixel_variance8x8 aom_highbd_10_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_10_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance16x16 aom_highbd_10_variance16x16_c
 
 unsigned int aom_highbd_10_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance16x32 aom_highbd_10_variance16x32_c
 
+unsigned int aom_highbd_10_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance16x4 aom_highbd_10_variance16x4_c
+
 unsigned int aom_highbd_10_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance16x8 aom_highbd_10_variance16x8_c
 
 unsigned int aom_highbd_10_variance2x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance2x2 aom_highbd_10_variance2x2_c
 
 unsigned int aom_highbd_10_variance2x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance2x4 aom_highbd_10_variance2x4_c
@@ -541,16 +811,22 @@ unsigned int aom_highbd_10_variance32x16
 #define aom_highbd_10_variance32x16 aom_highbd_10_variance32x16_c
 
 unsigned int aom_highbd_10_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance32x32 aom_highbd_10_variance32x32_c
 
 unsigned int aom_highbd_10_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance32x64 aom_highbd_10_variance32x64_c
 
+unsigned int aom_highbd_10_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance32x8 aom_highbd_10_variance32x8_c
+
+unsigned int aom_highbd_10_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance4x16 aom_highbd_10_variance4x16_c
+
 unsigned int aom_highbd_10_variance4x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance4x2 aom_highbd_10_variance4x2_c
 
 unsigned int aom_highbd_10_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance4x4 aom_highbd_10_variance4x4_c
 
 unsigned int aom_highbd_10_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance4x8 aom_highbd_10_variance4x8_c
@@ -559,16 +835,19 @@ unsigned int aom_highbd_10_variance64x32
 #define aom_highbd_10_variance64x32 aom_highbd_10_variance64x32_c
 
 unsigned int aom_highbd_10_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance64x64 aom_highbd_10_variance64x64_c
 
 unsigned int aom_highbd_10_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance8x16 aom_highbd_10_variance8x16_c
 
+unsigned int aom_highbd_10_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_10_variance8x32 aom_highbd_10_variance8x32_c
+
 unsigned int aom_highbd_10_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance8x4 aom_highbd_10_variance8x4_c
 
 unsigned int aom_highbd_10_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_10_variance8x8 aom_highbd_10_variance8x8_c
 
 void aom_highbd_12_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
 #define aom_highbd_12_get16x16var aom_highbd_12_get16x16var_c
@@ -577,43 +856,55 @@ void aom_highbd_12_get8x8var_c(const uin
 #define aom_highbd_12_get8x8var aom_highbd_12_get8x8var_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance16x16 aom_highbd_12_masked_sub_pixel_variance16x16_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance16x32 aom_highbd_12_masked_sub_pixel_variance16x32_c
 
+unsigned int aom_highbd_12_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance16x4 aom_highbd_12_masked_sub_pixel_variance16x4_c
+
 unsigned int aom_highbd_12_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance16x8 aom_highbd_12_masked_sub_pixel_variance16x8_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance32x16 aom_highbd_12_masked_sub_pixel_variance32x16_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance32x32 aom_highbd_12_masked_sub_pixel_variance32x32_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance32x64 aom_highbd_12_masked_sub_pixel_variance32x64_c
 
+unsigned int aom_highbd_12_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance32x8 aom_highbd_12_masked_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_12_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance4x16 aom_highbd_12_masked_sub_pixel_variance4x16_c
+
 unsigned int aom_highbd_12_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance4x4 aom_highbd_12_masked_sub_pixel_variance4x4_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance4x8 aom_highbd_12_masked_sub_pixel_variance4x8_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance64x32 aom_highbd_12_masked_sub_pixel_variance64x32_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance64x64 aom_highbd_12_masked_sub_pixel_variance64x64_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance8x16 aom_highbd_12_masked_sub_pixel_variance8x16_c
 
+unsigned int aom_highbd_12_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_12_masked_sub_pixel_variance8x32 aom_highbd_12_masked_sub_pixel_variance8x32_c
+
 unsigned int aom_highbd_12_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance8x4 aom_highbd_12_masked_sub_pixel_variance8x4_c
 
 unsigned int aom_highbd_12_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_12_masked_sub_pixel_variance8x8 aom_highbd_12_masked_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_12_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define aom_highbd_12_mse16x16 aom_highbd_12_mse16x16_c
@@ -628,172 +919,223 @@ unsigned int aom_highbd_12_mse8x8_c(cons
 #define aom_highbd_12_mse8x8 aom_highbd_12_mse8x8_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance16x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance16x16 aom_highbd_12_obmc_sub_pixel_variance16x16_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance16x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance16x32 aom_highbd_12_obmc_sub_pixel_variance16x32_c
 
+unsigned int aom_highbd_12_obmc_sub_pixel_variance16x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance16x4 aom_highbd_12_obmc_sub_pixel_variance16x4_c
+
 unsigned int aom_highbd_12_obmc_sub_pixel_variance16x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance16x8 aom_highbd_12_obmc_sub_pixel_variance16x8_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance32x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance32x16 aom_highbd_12_obmc_sub_pixel_variance32x16_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance32x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance32x32 aom_highbd_12_obmc_sub_pixel_variance32x32_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance32x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance32x64 aom_highbd_12_obmc_sub_pixel_variance32x64_c
 
+unsigned int aom_highbd_12_obmc_sub_pixel_variance32x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance32x8 aom_highbd_12_obmc_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_12_obmc_sub_pixel_variance4x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance4x16 aom_highbd_12_obmc_sub_pixel_variance4x16_c
+
 unsigned int aom_highbd_12_obmc_sub_pixel_variance4x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance4x4 aom_highbd_12_obmc_sub_pixel_variance4x4_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance4x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance4x8 aom_highbd_12_obmc_sub_pixel_variance4x8_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance64x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance64x32 aom_highbd_12_obmc_sub_pixel_variance64x32_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance64x64_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance64x64 aom_highbd_12_obmc_sub_pixel_variance64x64_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance8x16_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance8x16 aom_highbd_12_obmc_sub_pixel_variance8x16_c
 
+unsigned int aom_highbd_12_obmc_sub_pixel_variance8x32_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_sub_pixel_variance8x32 aom_highbd_12_obmc_sub_pixel_variance8x32_c
+
 unsigned int aom_highbd_12_obmc_sub_pixel_variance8x4_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance8x4 aom_highbd_12_obmc_sub_pixel_variance8x4_c
 
 unsigned int aom_highbd_12_obmc_sub_pixel_variance8x8_c(const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_sub_pixel_variance8x8 aom_highbd_12_obmc_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_12_obmc_variance16x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance16x16 aom_highbd_12_obmc_variance16x16_c
 
 unsigned int aom_highbd_12_obmc_variance16x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance16x32 aom_highbd_12_obmc_variance16x32_c
 
+unsigned int aom_highbd_12_obmc_variance16x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance16x4 aom_highbd_12_obmc_variance16x4_c
+
 unsigned int aom_highbd_12_obmc_variance16x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance16x8 aom_highbd_12_obmc_variance16x8_c
 
 unsigned int aom_highbd_12_obmc_variance32x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance32x16 aom_highbd_12_obmc_variance32x16_c
 
 unsigned int aom_highbd_12_obmc_variance32x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance32x32 aom_highbd_12_obmc_variance32x32_c
 
 unsigned int aom_highbd_12_obmc_variance32x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance32x64 aom_highbd_12_obmc_variance32x64_c
 
+unsigned int aom_highbd_12_obmc_variance32x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance32x8 aom_highbd_12_obmc_variance32x8_c
+
+unsigned int aom_highbd_12_obmc_variance4x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance4x16 aom_highbd_12_obmc_variance4x16_c
+
 unsigned int aom_highbd_12_obmc_variance4x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance4x4 aom_highbd_12_obmc_variance4x4_c
 
 unsigned int aom_highbd_12_obmc_variance4x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance4x8 aom_highbd_12_obmc_variance4x8_c
 
 unsigned int aom_highbd_12_obmc_variance64x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance64x32 aom_highbd_12_obmc_variance64x32_c
 
 unsigned int aom_highbd_12_obmc_variance64x64_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance64x64 aom_highbd_12_obmc_variance64x64_c
 
 unsigned int aom_highbd_12_obmc_variance8x16_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance8x16 aom_highbd_12_obmc_variance8x16_c
 
+unsigned int aom_highbd_12_obmc_variance8x32_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
+#define aom_highbd_12_obmc_variance8x32 aom_highbd_12_obmc_variance8x32_c
+
 unsigned int aom_highbd_12_obmc_variance8x4_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance8x4 aom_highbd_12_obmc_variance8x4_c
 
 unsigned int aom_highbd_12_obmc_variance8x8_c(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 #define aom_highbd_12_obmc_variance8x8 aom_highbd_12_obmc_variance8x8_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance16x16 aom_highbd_12_sub_pixel_avg_variance16x16_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance16x32 aom_highbd_12_sub_pixel_avg_variance16x32_c
 
+uint32_t aom_highbd_12_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance16x4 aom_highbd_12_sub_pixel_avg_variance16x4_c
+
 uint32_t aom_highbd_12_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance16x8 aom_highbd_12_sub_pixel_avg_variance16x8_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance32x16 aom_highbd_12_sub_pixel_avg_variance32x16_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance32x32 aom_highbd_12_sub_pixel_avg_variance32x32_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance32x64 aom_highbd_12_sub_pixel_avg_variance32x64_c
 
+uint32_t aom_highbd_12_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance32x8 aom_highbd_12_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance4x16 aom_highbd_12_sub_pixel_avg_variance4x16_c
+
 uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance4x4 aom_highbd_12_sub_pixel_avg_variance4x4_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance4x8 aom_highbd_12_sub_pixel_avg_variance4x8_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance64x32 aom_highbd_12_sub_pixel_avg_variance64x32_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance64x64 aom_highbd_12_sub_pixel_avg_variance64x64_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance8x16 aom_highbd_12_sub_pixel_avg_variance8x16_c
 
+uint32_t aom_highbd_12_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_12_sub_pixel_avg_variance8x32 aom_highbd_12_sub_pixel_avg_variance8x32_c
+
 uint32_t aom_highbd_12_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance8x4 aom_highbd_12_sub_pixel_avg_variance8x4_c
 
 uint32_t aom_highbd_12_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_12_sub_pixel_avg_variance8x8 aom_highbd_12_sub_pixel_avg_variance8x8_c
 
 uint32_t aom_highbd_12_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance16x16 aom_highbd_12_sub_pixel_variance16x16_c
 
 uint32_t aom_highbd_12_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance16x32 aom_highbd_12_sub_pixel_variance16x32_c
 
+uint32_t aom_highbd_12_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance16x4 aom_highbd_12_sub_pixel_variance16x4_c
+
 uint32_t aom_highbd_12_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance16x8 aom_highbd_12_sub_pixel_variance16x8_c
 
 uint32_t aom_highbd_12_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance32x16 aom_highbd_12_sub_pixel_variance32x16_c
 
 uint32_t aom_highbd_12_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance32x32 aom_highbd_12_sub_pixel_variance32x32_c
 
 uint32_t aom_highbd_12_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance32x64 aom_highbd_12_sub_pixel_variance32x64_c
 
+uint32_t aom_highbd_12_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance32x8 aom_highbd_12_sub_pixel_variance32x8_c
+
+uint32_t aom_highbd_12_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance4x16 aom_highbd_12_sub_pixel_variance4x16_c
+
 uint32_t aom_highbd_12_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance4x4 aom_highbd_12_sub_pixel_variance4x4_c
 
 uint32_t aom_highbd_12_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance4x8 aom_highbd_12_sub_pixel_variance4x8_c
 
 uint32_t aom_highbd_12_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance64x32 aom_highbd_12_sub_pixel_variance64x32_c
 
 uint32_t aom_highbd_12_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance64x64 aom_highbd_12_sub_pixel_variance64x64_c
 
 uint32_t aom_highbd_12_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance8x16 aom_highbd_12_sub_pixel_variance8x16_c
 
+uint32_t aom_highbd_12_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_sub_pixel_variance8x32 aom_highbd_12_sub_pixel_variance8x32_c
+
 uint32_t aom_highbd_12_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance8x4 aom_highbd_12_sub_pixel_variance8x4_c
 
 uint32_t aom_highbd_12_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_12_sub_pixel_variance8x8 aom_highbd_12_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_12_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance16x16 aom_highbd_12_variance16x16_c
 
 unsigned int aom_highbd_12_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance16x32 aom_highbd_12_variance16x32_c
 
+unsigned int aom_highbd_12_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance16x4 aom_highbd_12_variance16x4_c
+
 unsigned int aom_highbd_12_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance16x8 aom_highbd_12_variance16x8_c
 
 unsigned int aom_highbd_12_variance2x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance2x2 aom_highbd_12_variance2x2_c
 
 unsigned int aom_highbd_12_variance2x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance2x4 aom_highbd_12_variance2x4_c
@@ -802,16 +1144,22 @@ unsigned int aom_highbd_12_variance32x16
 #define aom_highbd_12_variance32x16 aom_highbd_12_variance32x16_c
 
 unsigned int aom_highbd_12_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance32x32 aom_highbd_12_variance32x32_c
 
 unsigned int aom_highbd_12_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance32x64 aom_highbd_12_variance32x64_c
 
+unsigned int aom_highbd_12_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance32x8 aom_highbd_12_variance32x8_c
+
+unsigned int aom_highbd_12_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance4x16 aom_highbd_12_variance4x16_c
+
 unsigned int aom_highbd_12_variance4x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance4x2 aom_highbd_12_variance4x2_c
 
 unsigned int aom_highbd_12_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance4x4 aom_highbd_12_variance4x4_c
 
 unsigned int aom_highbd_12_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance4x8 aom_highbd_12_variance4x8_c
@@ -820,16 +1168,19 @@ unsigned int aom_highbd_12_variance64x32
 #define aom_highbd_12_variance64x32 aom_highbd_12_variance64x32_c
 
 unsigned int aom_highbd_12_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance64x64 aom_highbd_12_variance64x64_c
 
 unsigned int aom_highbd_12_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance8x16 aom_highbd_12_variance8x16_c
 
+unsigned int aom_highbd_12_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_12_variance8x32 aom_highbd_12_variance8x32_c
+
 unsigned int aom_highbd_12_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance8x4 aom_highbd_12_variance8x4_c
 
 unsigned int aom_highbd_12_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_12_variance8x8 aom_highbd_12_variance8x8_c
 
 void aom_highbd_8_get16x16var_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
 #define aom_highbd_8_get16x16var aom_highbd_8_get16x16var_c
@@ -838,43 +1189,55 @@ void aom_highbd_8_get8x8var_c(const uint
 #define aom_highbd_8_get8x8var aom_highbd_8_get8x8var_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance16x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance16x16 aom_highbd_8_masked_sub_pixel_variance16x16_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance16x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance16x32 aom_highbd_8_masked_sub_pixel_variance16x32_c
 
+unsigned int aom_highbd_8_masked_sub_pixel_variance16x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance16x4 aom_highbd_8_masked_sub_pixel_variance16x4_c
+
 unsigned int aom_highbd_8_masked_sub_pixel_variance16x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance16x8 aom_highbd_8_masked_sub_pixel_variance16x8_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance32x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance32x16 aom_highbd_8_masked_sub_pixel_variance32x16_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance32x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance32x32 aom_highbd_8_masked_sub_pixel_variance32x32_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance32x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance32x64 aom_highbd_8_masked_sub_pixel_variance32x64_c
 
+unsigned int aom_highbd_8_masked_sub_pixel_variance32x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance32x8 aom_highbd_8_masked_sub_pixel_variance32x8_c
+
+unsigned int aom_highbd_8_masked_sub_pixel_variance4x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance4x16 aom_highbd_8_masked_sub_pixel_variance4x16_c
+
 unsigned int aom_highbd_8_masked_sub_pixel_variance4x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance4x4 aom_highbd_8_masked_sub_pixel_variance4x4_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance4x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance4x8 aom_highbd_8_masked_sub_pixel_variance4x8_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance64x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance64x32 aom_highbd_8_masked_sub_pixel_variance64x32_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance64x64_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance64x64 aom_highbd_8_masked_sub_pixel_variance64x64_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance8x16_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance8x16 aom_highbd_8_masked_sub_pixel_variance8x16_c
 
+unsigned int aom_highbd_8_masked_sub_pixel_variance8x32_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
+#define aom_highbd_8_masked_sub_pixel_variance8x32 aom_highbd_8_masked_sub_pixel_variance8x32_c
+
 unsigned int aom_highbd_8_masked_sub_pixel_variance8x4_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance8x4 aom_highbd_8_masked_sub_pixel_variance8x4_c
 
 unsigned int aom_highbd_8_masked_sub_pixel_variance8x8_c(const uint8_t *src, int src_stride, int xoffset, int yoffset, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask, unsigned int *sse);
 #define aom_highbd_8_masked_sub_pixel_variance8x8 aom_highbd_8_masked_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_8_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define aom_highbd_8_mse16x16 aom_highbd_8_mse16x16_c
@@ -889,94 +1252,121 @@ unsigned int aom_highbd_8_mse8x8_c(const
 #define aom_highbd_8_mse8x8 aom_highbd_8_mse8x8_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance16x16 aom_highbd_8_sub_pixel_avg_variance16x16_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance16x32 aom_highbd_8_sub_pixel_avg_variance16x32_c
 
+uint32_t aom_highbd_8_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance16x4 aom_highbd_8_sub_pixel_avg_variance16x4_c
+
 uint32_t aom_highbd_8_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance16x8 aom_highbd_8_sub_pixel_avg_variance16x8_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance32x16 aom_highbd_8_sub_pixel_avg_variance32x16_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance32x32 aom_highbd_8_sub_pixel_avg_variance32x32_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance32x64 aom_highbd_8_sub_pixel_avg_variance32x64_c
 
+uint32_t aom_highbd_8_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance32x8 aom_highbd_8_sub_pixel_avg_variance32x8_c
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance4x16 aom_highbd_8_sub_pixel_avg_variance4x16_c
+
 uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance4x4 aom_highbd_8_sub_pixel_avg_variance4x4_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance4x8 aom_highbd_8_sub_pixel_avg_variance4x8_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance64x32 aom_highbd_8_sub_pixel_avg_variance64x32_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance64x64 aom_highbd_8_sub_pixel_avg_variance64x64_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance8x16 aom_highbd_8_sub_pixel_avg_variance8x16_c
 
+uint32_t aom_highbd_8_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
+#define aom_highbd_8_sub_pixel_avg_variance8x32 aom_highbd_8_sub_pixel_avg_variance8x32_c
+
 uint32_t aom_highbd_8_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance8x4 aom_highbd_8_sub_pixel_avg_variance8x4_c
 
 uint32_t aom_highbd_8_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 #define aom_highbd_8_sub_pixel_avg_variance8x8 aom_highbd_8_sub_pixel_avg_variance8x8_c
 
 uint32_t aom_highbd_8_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance16x16 aom_highbd_8_sub_pixel_variance16x16_c
 
 uint32_t aom_highbd_8_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance16x32 aom_highbd_8_sub_pixel_variance16x32_c
 
+uint32_t aom_highbd_8_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance16x4 aom_highbd_8_sub_pixel_variance16x4_c
+
 uint32_t aom_highbd_8_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance16x8 aom_highbd_8_sub_pixel_variance16x8_c
 
 uint32_t aom_highbd_8_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance32x16 aom_highbd_8_sub_pixel_variance32x16_c
 
 uint32_t aom_highbd_8_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance32x32 aom_highbd_8_sub_pixel_variance32x32_c
 
 uint32_t aom_highbd_8_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance32x64 aom_highbd_8_sub_pixel_variance32x64_c
 
+uint32_t aom_highbd_8_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance32x8 aom_highbd_8_sub_pixel_variance32x8_c
+
+uint32_t aom_highbd_8_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance4x16 aom_highbd_8_sub_pixel_variance4x16_c
+
 uint32_t aom_highbd_8_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance4x4 aom_highbd_8_sub_pixel_variance4x4_c
 
 uint32_t aom_highbd_8_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance4x8 aom_highbd_8_sub_pixel_variance4x8_c
 
 uint32_t aom_highbd_8_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance64x32 aom_highbd_8_sub_pixel_variance64x32_c
 
 uint32_t aom_highbd_8_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance64x64 aom_highbd_8_sub_pixel_variance64x64_c
 
 uint32_t aom_highbd_8_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance8x16 aom_highbd_8_sub_pixel_variance8x16_c
 
+uint32_t aom_highbd_8_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_sub_pixel_variance8x32 aom_highbd_8_sub_pixel_variance8x32_c
+
 uint32_t aom_highbd_8_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance8x4 aom_highbd_8_sub_pixel_variance8x4_c
 
 uint32_t aom_highbd_8_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 #define aom_highbd_8_sub_pixel_variance8x8 aom_highbd_8_sub_pixel_variance8x8_c
 
 unsigned int aom_highbd_8_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance16x16 aom_highbd_8_variance16x16_c
 
 unsigned int aom_highbd_8_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance16x32 aom_highbd_8_variance16x32_c
 
+unsigned int aom_highbd_8_variance16x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance16x4 aom_highbd_8_variance16x4_c
+
 unsigned int aom_highbd_8_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance16x8 aom_highbd_8_variance16x8_c
 
 unsigned int aom_highbd_8_variance2x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance2x2 aom_highbd_8_variance2x2_c
 
 unsigned int aom_highbd_8_variance2x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance2x4 aom_highbd_8_variance2x4_c
@@ -985,16 +1375,22 @@ unsigned int aom_highbd_8_variance32x16_
 #define aom_highbd_8_variance32x16 aom_highbd_8_variance32x16_c
 
 unsigned int aom_highbd_8_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance32x32 aom_highbd_8_variance32x32_c
 
 unsigned int aom_highbd_8_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance32x64 aom_highbd_8_variance32x64_c
 
+unsigned int aom_highbd_8_variance32x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance32x8 aom_highbd_8_variance32x8_c
+
+unsigned int aom_highbd_8_variance4x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance4x16 aom_highbd_8_variance4x16_c
+
 unsigned int aom_highbd_8_variance4x2_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance4x2 aom_highbd_8_variance4x2_c
 
 unsigned int aom_highbd_8_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance4x4 aom_highbd_8_variance4x4_c
 
 unsigned int aom_highbd_8_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance4x8 aom_highbd_8_variance4x8_c
@@ -1003,16 +1399,19 @@ unsigned int aom_highbd_8_variance64x32_
 #define aom_highbd_8_variance64x32 aom_highbd_8_variance64x32_c
 
 unsigned int aom_highbd_8_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance64x64 aom_highbd_8_variance64x64_c
 
 unsigned int aom_highbd_8_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance8x16 aom_highbd_8_variance8x16_c
 
+unsigned int aom_highbd_8_variance8x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
+#define aom_highbd_8_variance8x32 aom_highbd_8_variance8x32_c
+
 unsigned int aom_highbd_8_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance8x4 aom_highbd_8_variance8x4_c
 
 unsigned int aom_highbd_8_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
 #define aom_highbd_8_variance8x8 aom_highbd_8_variance8x8_c
 
 void aom_highbd_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd);
 #define aom_highbd_blend_a64_hmask aom_highbd_blend_a64_hmask_c
@@ -1021,23 +1420,23 @@ void aom_highbd_blend_a64_mask_c(uint8_t
 #define aom_highbd_blend_a64_mask aom_highbd_blend_a64_mask_c
 
 void aom_highbd_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w, int bd);
 #define aom_highbd_blend_a64_vmask aom_highbd_blend_a64_vmask_c
 
 void aom_highbd_comp_avg_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
 #define aom_highbd_comp_avg_pred aom_highbd_comp_avg_pred_c
 
-void aom_highbd_comp_avg_upsampled_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride);
+void aom_highbd_comp_avg_upsampled_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref8, int ref_stride, int bd);
 #define aom_highbd_comp_avg_upsampled_pred aom_highbd_comp_avg_upsampled_pred_c
 
 void aom_highbd_comp_mask_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 #define aom_highbd_comp_mask_pred aom_highbd_comp_mask_pred_c
 
-void aom_highbd_comp_mask_upsampled_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
+void aom_highbd_comp_mask_upsampled_pred_c(uint16_t *comp_pred, const uint8_t *pred8, int width, int height, int subsample_x_q3, int subsample_y_q3, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask, int bd);
 #define aom_highbd_comp_mask_upsampled_pred aom_highbd_comp_mask_upsampled_pred_c
 
 void aom_highbd_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
 #define aom_highbd_convolve8 aom_highbd_convolve8_c
 
 void aom_highbd_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
 #define aom_highbd_convolve8_avg aom_highbd_convolve8_avg_c
 
@@ -1057,223 +1456,391 @@ void aom_highbd_convolve_avg_c(const uin
 #define aom_highbd_convolve_avg aom_highbd_convolve_avg_c
 
 void aom_highbd_convolve_copy_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
 #define aom_highbd_convolve_copy aom_highbd_convolve_copy_c
 
 void aom_highbd_d117_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_16x16 aom_highbd_d117_predictor_16x16_c
 
+void aom_highbd_d117_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_16x32 aom_highbd_d117_predictor_16x32_c
+
+void aom_highbd_d117_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_16x8 aom_highbd_d117_predictor_16x8_c
+
 void aom_highbd_d117_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_2x2 aom_highbd_d117_predictor_2x2_c
 
+void aom_highbd_d117_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_32x16 aom_highbd_d117_predictor_32x16_c
+
 void aom_highbd_d117_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_32x32 aom_highbd_d117_predictor_32x32_c
 
 void aom_highbd_d117_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_4x4 aom_highbd_d117_predictor_4x4_c
 
+void aom_highbd_d117_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_4x8 aom_highbd_d117_predictor_4x8_c
+
+void aom_highbd_d117_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_8x16 aom_highbd_d117_predictor_8x16_c
+
+void aom_highbd_d117_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_8x4 aom_highbd_d117_predictor_8x4_c
+
 void aom_highbd_d117_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_8x8 aom_highbd_d117_predictor_8x8_c
 
 void aom_highbd_d135_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_16x16 aom_highbd_d135_predictor_16x16_c
 
+void aom_highbd_d135_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_16x32 aom_highbd_d135_predictor_16x32_c
+
+void aom_highbd_d135_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_16x8 aom_highbd_d135_predictor_16x8_c
+
 void aom_highbd_d135_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_2x2 aom_highbd_d135_predictor_2x2_c
 
+void aom_highbd_d135_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_32x16 aom_highbd_d135_predictor_32x16_c
+
 void aom_highbd_d135_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_32x32 aom_highbd_d135_predictor_32x32_c
 
 void aom_highbd_d135_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_4x4 aom_highbd_d135_predictor_4x4_c
 
+void aom_highbd_d135_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_4x8 aom_highbd_d135_predictor_4x8_c
+
+void aom_highbd_d135_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_8x16 aom_highbd_d135_predictor_8x16_c
+
+void aom_highbd_d135_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_8x4 aom_highbd_d135_predictor_8x4_c
+
 void aom_highbd_d135_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_8x8 aom_highbd_d135_predictor_8x8_c
 
 void aom_highbd_d153_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_16x16 aom_highbd_d153_predictor_16x16_c
 
+void aom_highbd_d153_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_16x32 aom_highbd_d153_predictor_16x32_c
+
+void aom_highbd_d153_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_16x8 aom_highbd_d153_predictor_16x8_c
+
 void aom_highbd_d153_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_2x2 aom_highbd_d153_predictor_2x2_c
 
+void aom_highbd_d153_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_32x16 aom_highbd_d153_predictor_32x16_c
+
 void aom_highbd_d153_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_32x32 aom_highbd_d153_predictor_32x32_c
 
 void aom_highbd_d153_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_4x4 aom_highbd_d153_predictor_4x4_c
 
+void aom_highbd_d153_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_4x8 aom_highbd_d153_predictor_4x8_c
+
+void aom_highbd_d153_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_8x16 aom_highbd_d153_predictor_8x16_c
+
+void aom_highbd_d153_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_8x4 aom_highbd_d153_predictor_8x4_c
+
 void aom_highbd_d153_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_8x8 aom_highbd_d153_predictor_8x8_c
 
 void aom_highbd_d207e_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_16x16 aom_highbd_d207e_predictor_16x16_c
 
+void aom_highbd_d207e_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_16x32 aom_highbd_d207e_predictor_16x32_c
+
+void aom_highbd_d207e_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_16x8 aom_highbd_d207e_predictor_16x8_c
+
 void aom_highbd_d207e_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_2x2 aom_highbd_d207e_predictor_2x2_c
 
+void aom_highbd_d207e_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_32x16 aom_highbd_d207e_predictor_32x16_c
+
 void aom_highbd_d207e_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_32x32 aom_highbd_d207e_predictor_32x32_c
 
 void aom_highbd_d207e_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_4x4 aom_highbd_d207e_predictor_4x4_c
 
+void aom_highbd_d207e_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_4x8 aom_highbd_d207e_predictor_4x8_c
+
+void aom_highbd_d207e_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_8x16 aom_highbd_d207e_predictor_8x16_c
+
+void aom_highbd_d207e_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d207e_predictor_8x4 aom_highbd_d207e_predictor_8x4_c
+
 void aom_highbd_d207e_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_8x8 aom_highbd_d207e_predictor_8x8_c
 
 void aom_highbd_d45e_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_16x16 aom_highbd_d45e_predictor_16x16_c
 
+void aom_highbd_d45e_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_16x32 aom_highbd_d45e_predictor_16x32_c
+
+void aom_highbd_d45e_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_16x8 aom_highbd_d45e_predictor_16x8_c
+
 void aom_highbd_d45e_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_2x2 aom_highbd_d45e_predictor_2x2_c
 
+void aom_highbd_d45e_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_32x16 aom_highbd_d45e_predictor_32x16_c
+
 void aom_highbd_d45e_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_32x32 aom_highbd_d45e_predictor_32x32_c
 
 void aom_highbd_d45e_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_4x4 aom_highbd_d45e_predictor_4x4_c
 
+void aom_highbd_d45e_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_4x8 aom_highbd_d45e_predictor_4x8_c
+
+void aom_highbd_d45e_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_8x16 aom_highbd_d45e_predictor_8x16_c
+
+void aom_highbd_d45e_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_8x4 aom_highbd_d45e_predictor_8x4_c
+
 void aom_highbd_d45e_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_8x8 aom_highbd_d45e_predictor_8x8_c
 
 void aom_highbd_d63e_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_16x16 aom_highbd_d63e_predictor_16x16_c
 
+void aom_highbd_d63e_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_16x32 aom_highbd_d63e_predictor_16x32_c
+
+void aom_highbd_d63e_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_16x8 aom_highbd_d63e_predictor_16x8_c
+
 void aom_highbd_d63e_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_2x2 aom_highbd_d63e_predictor_2x2_c
 
+void aom_highbd_d63e_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_32x16 aom_highbd_d63e_predictor_32x16_c
+
 void aom_highbd_d63e_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_32x32 aom_highbd_d63e_predictor_32x32_c
 
 void aom_highbd_d63e_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_4x4 aom_highbd_d63e_predictor_4x4_c
 
+void aom_highbd_d63e_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_4x8 aom_highbd_d63e_predictor_4x8_c
+
+void aom_highbd_d63e_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_8x16 aom_highbd_d63e_predictor_8x16_c
+
+void aom_highbd_d63e_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d63e_predictor_8x4 aom_highbd_d63e_predictor_8x4_c
+
 void aom_highbd_d63e_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_8x8 aom_highbd_d63e_predictor_8x8_c
 
 void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_16x16 aom_highbd_dc_128_predictor_16x16_c
 
+void aom_highbd_dc_128_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x32 aom_highbd_dc_128_predictor_16x32_c
+
+void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_c
+
 void aom_highbd_dc_128_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_c
 
+void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_c
+
 void aom_highbd_dc_128_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_32x32 aom_highbd_dc_128_predictor_32x32_c
 
 void aom_highbd_dc_128_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_4x4 aom_highbd_dc_128_predictor_4x4_c
 
+void aom_highbd_dc_128_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x8 aom_highbd_dc_128_predictor_4x8_c
+
+void aom_highbd_dc_128_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x16 aom_highbd_dc_128_predictor_8x16_c
+
+void aom_highbd_dc_128_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x4 aom_highbd_dc_128_predictor_8x4_c
+
 void aom_highbd_dc_128_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_8x8 aom_highbd_dc_128_predictor_8x8_c
 
 void aom_highbd_dc_left_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_16x16 aom_highbd_dc_left_predictor_16x16_c
 
+void aom_highbd_dc_left_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x32 aom_highbd_dc_left_predictor_16x32_c
+
+void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_c
+
 void aom_highbd_dc_left_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_c
 
+void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_c
+
 void aom_highbd_dc_left_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_32x32 aom_highbd_dc_left_predictor_32x32_c
 
 void aom_highbd_dc_left_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_4x4 aom_highbd_dc_left_predictor_4x4_c
 
+void aom_highbd_dc_left_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x8 aom_highbd_dc_left_predictor_4x8_c
+
+void aom_highbd_dc_left_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x16 aom_highbd_dc_left_predictor_8x16_c
+
+void aom_highbd_dc_left_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x4 aom_highbd_dc_left_predictor_8x4_c
+
 void aom_highbd_dc_left_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_8x8 aom_highbd_dc_left_predictor_8x8_c
 
 void aom_highbd_dc_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_16x16 aom_highbd_dc_predictor_16x16_c
 
+void aom_highbd_dc_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x32 aom_highbd_dc_predictor_16x32_c
+
+void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_c
+
 void aom_highbd_dc_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_c
 
+void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_c
+
 void aom_highbd_dc_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_32x32 aom_highbd_dc_predictor_32x32_c
 
 void aom_highbd_dc_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_4x4 aom_highbd_dc_predictor_4x4_c
 
+void aom_highbd_dc_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x8 aom_highbd_dc_predictor_4x8_c
+
+void aom_highbd_dc_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x16 aom_highbd_dc_predictor_8x16_c
+
+void aom_highbd_dc_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x4 aom_highbd_dc_predictor_8x4_c
+
 void aom_highbd_dc_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_8x8 aom_highbd_dc_predictor_8x8_c
 
 void aom_highbd_dc_top_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_16x16 aom_highbd_dc_top_predictor_16x16_c
 
+void aom_highbd_dc_top_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x32 aom_highbd_dc_top_predictor_16x32_c
+
+void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_c
+
 void aom_highbd_dc_top_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_c
 
+void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_c
+
 void aom_highbd_dc_top_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_32x32 aom_highbd_dc_top_predictor_32x32_c
 
 void aom_highbd_dc_top_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_4x4 aom_highbd_dc_top_predictor_4x4_c
 
+void aom_highbd_dc_top_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x8 aom_highbd_dc_top_predictor_4x8_c
+
+void aom_highbd_dc_top_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x16 aom_highbd_dc_top_predictor_8x16_c
+
+void aom_highbd_dc_top_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);