Turn on libvpx x86 optimizations [DO NOT MERGE]

With yasm available we can use the assembly:
https://android-review.googlesource.com/#/c/96972/

Comparing with the VP9 video playback with C-reference software decode,
before and after these patches applied,

Performance data with 2 test clips:
Clips                                    bit rate    fps(after)  fps(before)
vp9_720p_qp10to20_1510Kbps_30fps.webm	  1510Kbps      30	     12
vp9_720p_qp20to30_840Kbps_30fps.webm	  840Kbps       30	     13

(cherry-picked from afecaee6099284387688b9ce17ec3642ec4e5c1b)

Change-Id: I9c67e5ae302f923134e073f4bafb09fb1f3fe94d

9 files changed, 1391 insertions, 2 deletions

diff --git a/config.x86.mk b/config.x86.mk
index 055a793..5d604b9 100644
--- a/config.x86.mk
+++ b/config.x86.mk
@@ -3,7 +3,7 @@
 # libvpx_codec_srcs_c_x86
 # libvpx_codec_srcs_asm_x86
 
-libvpx_target := generic
+libvpx_target := x86
 
 libvpx_config_dir_x86 := $(LOCAL_PATH)/$(libvpx_target)
 libvpx_codec_srcs := $(sort $(shell cat $(libvpx_config_dir_x86)/libvpx_srcs.txt))
@@ -13,4 +13,5 @@ libvpx_codec_srcs_c_x86 := $(addprefix libvpx/, $(filter-out vpx_config.c, \
     $(filter %.c, $(libvpx_codec_srcs)))) \
     $(libvpx_target)/vpx_config.c
 
-libvpx_codec_srcs_asm_x86 := $(filter %.asm.s, $(libvpx_codec_srcs))
+# X86 asm files are processed by the system and sent to yasm
+libvpx_codec_srcs_c_x86 += $(addprefix libvpx/, $(filter %.asm, $(libvpx_codec_srcs)))
diff --git a/x86/libvpx_srcs.txt b/x86/libvpx_srcs.txt
new file mode 100644
index 0000000..0c59ce9
--- /dev/null
+++ b/x86/libvpx_srcs.txt
@@ -0,0 +1,299 @@
+build/make/rtcd.pl
+build/make/version.sh
+CHANGELOG
+libs.mk
+third_party/x86inc/x86inc.asm
+vp8/common/alloccommon.c
+vp8/common/alloccommon.h
+vp8/common/blockd.c
+vp8/common/blockd.h
+vp8/common/coefupdateprobs.h
+vp8/common/common.h
+vp8/common/debugmodes.c
+vp8/common/default_coef_probs.h
+vp8/common/dequantize.c
+vp8/common/entropy.c
+vp8/common/entropy.h
+vp8/common/entropymode.c
+vp8/common/entropymode.h
+vp8/common/entropymv.c
+vp8/common/entropymv.h
+vp8/common/extend.c
+vp8/common/extend.h
+vp8/common/filter.c
+vp8/common/filter.h
+vp8/common/findnearmv.c
+vp8/common/findnearmv.h
+vp8/common/generic/systemdependent.c
+vp8/common/header.h
+vp8/common/idct_blk.c
+vp8/common/idctllm.c
+vp8/common/invtrans.h
+vp8/common/loopfilter.c
+vp8/common/loopfilter_filters.c
+vp8/common/loopfilter.h
+vp8/common/mbpitch.c
+vp8/common/mfqe.c
+vp8/common/modecont.c
+vp8/common/modecont.h
+vp8/common/mv.h
+vp8/common/onyxc_int.h
+vp8/common/onyxd.h
+vp8/common/onyx.h
+vp8/common/postproc.c
+vp8/common/postproc.h
+vp8/common/ppflags.h
+vp8/common/pragmas.h
+vp8/common/quant_common.c
+vp8/common/quant_common.h
+vp8/common/reconinter.c
+vp8/common/reconinter.h
+vp8/common/reconintra4x4.c
+vp8/common/reconintra4x4.h
+vp8/common/reconintra.c
+vp8/common/rtcd.c
+vp8/common/rtcd_defs.pl
+vp8/common/sad_c.c
+vp8/common/setupintrarecon.c
+vp8/common/setupintrarecon.h
+vp8/common/swapyv12buffer.c
+vp8/common/swapyv12buffer.h
+vp8/common/systemdependent.h
+vp8/common/threading.h
+vp8/common/treecoder.c
+vp8/common/treecoder.h
+vp8/common/variance_c.c
+vp8/common/variance.h
+vp8/common/vp8_entropymodedata.h
+vp8/common/x86/dequantize_mmx.asm
+vp8/common/x86/filter_x86.c
+vp8/common/x86/filter_x86.h
+vp8/common/x86/idct_blk_mmx.c
+vp8/common/x86/idct_blk_sse2.c
+vp8/common/x86/idctllm_mmx.asm
+vp8/common/x86/idctllm_sse2.asm
+vp8/common/x86/iwalsh_mmx.asm
+vp8/common/x86/iwalsh_sse2.asm
+vp8/common/x86/loopfilter_mmx.asm
+vp8/common/x86/loopfilter_sse2.asm
+vp8/common/x86/loopfilter_x86.c
+vp8/common/x86/mfqe_sse2.asm
+vp8/common/x86/postproc_mmx.asm
+vp8/common/x86/postproc_sse2.asm
+vp8/common/x86/postproc_x86.c
+vp8/common/x86/recon_mmx.asm
+vp8/common/x86/recon_sse2.asm
+vp8/common/x86/recon_wrapper_sse2.c
+vp8/common/x86/sad_mmx.asm
+vp8/common/x86/sad_sse2.asm
+vp8/common/x86/subpixel_mmx.asm
+vp8/common/x86/subpixel_sse2.asm
+vp8/common/x86/variance_impl_mmx.asm
+vp8/common/x86/variance_impl_sse2.asm
+vp8/common/x86/variance_mmx.c
+vp8/common/x86/variance_sse2.c
+vp8/common/x86/vp8_asm_stubs.c
+vp8/decoder/dboolhuff.c
+vp8/decoder/dboolhuff.h
+vp8/decoder/decodeframe.c
+vp8/decoder/decodemv.c
+vp8/decoder/decodemv.h
+vp8/decoder/decoderthreading.h
+vp8/decoder/detokenize.c
+vp8/decoder/detokenize.h
+vp8/decoder/onyxd_if.c
+vp8/decoder/onyxd_int.h
+vp8/decoder/threading.c
+vp8/decoder/treereader.h
+vp8/encoder/bitstream.c
+vp8/encoder/bitstream.h
+vp8/encoder/block.h
+vp8/encoder/boolhuff.c
+vp8/encoder/boolhuff.h
+vp8/encoder/dct.c
+vp8/encoder/dct_value_cost.h
+vp8/encoder/dct_value_tokens.h
+vp8/encoder/defaultcoefcounts.h
+vp8/encoder/denoising.c
+vp8/encoder/denoising.h
+vp8/encoder/encodeframe.c
+vp8/encoder/encodeframe.h
+vp8/encoder/encodeintra.c
+vp8/encoder/encodeintra.h
+vp8/encoder/encodemb.c
+vp8/encoder/encodemb.h
+vp8/encoder/encodemv.c
+vp8/encoder/encodemv.h
+vp8/encoder/ethreading.c
+vp8/encoder/firstpass.c
+vp8/encoder/firstpass.h
+vp8/encoder/lookahead.c
+vp8/encoder/lookahead.h
+vp8/encoder/mcomp.c
+vp8/encoder/mcomp.h
+vp8/encoder/modecosts.c
+vp8/encoder/modecosts.h
+vp8/encoder/onyx_if.c
+vp8/encoder/onyx_int.h
+vp8/encoder/pickinter.c
+vp8/encoder/pickinter.h
+vp8/encoder/picklpf.c
+vp8/encoder/quantize.c
+vp8/encoder/quantize.h
+vp8/encoder/ratectrl.c
+vp8/encoder/ratectrl.h
+vp8/encoder/rdopt.c
+vp8/encoder/rdopt.h
+vp8/encoder/segmentation.c
+vp8/encoder/segmentation.h
+vp8/encoder/temporal_filter.c
+vp8/encoder/tokenize.c
+vp8/encoder/tokenize.h
+vp8/encoder/treewriter.c
+vp8/encoder/treewriter.h
+vp8/encoder/vp8_asm_enc_offsets.c
+vp8/encoder/x86/dct_mmx.asm
+vp8/encoder/x86/dct_sse2.asm
+vp8/encoder/x86/denoising_sse2.c
+vp8/encoder/x86/encodeopt.asm
+vp8/encoder/x86/fwalsh_sse2.asm
+vp8/encoder/x86/quantize_mmx.asm
+vp8/encoder/x86/quantize_sse2.c
+vp8/encoder/x86/subtract_mmx.asm
+vp8/encoder/x86/subtract_sse2.asm
+vp8/encoder/x86/temporal_filter_apply_sse2.asm
+vp8/encoder/x86/vp8_enc_stubs_mmx.c
+vp8/encoder/x86/vp8_enc_stubs_sse2.c
+vp8/vp8_common.mk
+vp8/vp8_cx_iface.c
+vp8/vp8cx.mk
+vp8/vp8_dx_iface.c
+vp8/vp8dx.mk
+vp9/common/vp9_alloccommon.c
+vp9/common/vp9_alloccommon.h
+vp9/common/vp9_blockd.c
+vp9/common/vp9_blockd.h
+vp9/common/vp9_common_data.c
+vp9/common/vp9_common_data.h
+vp9/common/vp9_common.h
+vp9/common/vp9_convolve.c
+vp9/common/vp9_convolve.h
+vp9/common/vp9_debugmodes.c
+vp9/common/vp9_entropy.c
+vp9/common/vp9_entropy.h
+vp9/common/vp9_entropymode.c
+vp9/common/vp9_entropymode.h
+vp9/common/vp9_entropymv.c
+vp9/common/vp9_entropymv.h
+vp9/common/vp9_enums.h
+vp9/common/vp9_filter.c
+vp9/common/vp9_filter.h
+vp9/common/vp9_frame_buffers.c
+vp9/common/vp9_frame_buffers.h
+vp9/common/vp9_idct.c
+vp9/common/vp9_idct.h
+vp9/common/vp9_loopfilter.c
+vp9/common/vp9_loopfilter_filters.c
+vp9/common/vp9_loopfilter.h
+vp9/common/vp9_mv.h
+vp9/common/vp9_mvref_common.c
+vp9/common/vp9_mvref_common.h
+vp9/common/vp9_onyxc_int.h
+vp9/common/vp9_ppflags.h
+vp9/common/vp9_pragmas.h
+vp9/common/vp9_pred_common.c
+vp9/common/vp9_pred_common.h
+vp9/common/vp9_prob.c
+vp9/common/vp9_prob.h
+vp9/common/vp9_quant_common.c
+vp9/common/vp9_quant_common.h
+vp9/common/vp9_reconinter.c
+vp9/common/vp9_reconinter.h
+vp9/common/vp9_reconintra.c
+vp9/common/vp9_reconintra.h
+vp9/common/vp9_rtcd.c
+vp9/common/vp9_rtcd_defs.pl
+vp9/common/vp9_scale.c
+vp9/common/vp9_scale.h
+vp9/common/vp9_scan.c
+vp9/common/vp9_scan.h
+vp9/common/vp9_seg_common.c
+vp9/common/vp9_seg_common.h
+vp9/common/vp9_systemdependent.h
+vp9/common/vp9_textblit.h
+vp9/common/vp9_tile_common.c
+vp9/common/vp9_tile_common.h
+vp9/common/x86/vp9_asm_stubs.c
+vp9/common/x86/vp9_copy_sse2.asm
+vp9/common/x86/vp9_idct_intrin_sse2.c
+vp9/common/x86/vp9_intrapred_sse2.asm
+vp9/common/x86/vp9_loopfilter_intrin_sse2.c
+vp9/common/x86/vp9_loopfilter_mmx.asm
+vp9/common/x86/vp9_postproc_x86.h
+vp9/common/x86/vp9_subpixel_8t_sse2.asm
+vp9/common/x86/vp9_subpixel_bilinear_sse2.asm
+vp9/decoder/vp9_decodeframe.c
+vp9/decoder/vp9_decodeframe.h
+vp9/decoder/vp9_decodemv.c
+vp9/decoder/vp9_decodemv.h
+vp9/decoder/vp9_decoder.c
+vp9/decoder/vp9_decoder.h
+vp9/decoder/vp9_detokenize.c
+vp9/decoder/vp9_detokenize.h
+vp9/decoder/vp9_dsubexp.c
+vp9/decoder/vp9_dsubexp.h
+vp9/decoder/vp9_dthread.c
+vp9/decoder/vp9_dthread.h
+vp9/decoder/vp9_read_bit_buffer.c
+vp9/decoder/vp9_read_bit_buffer.h
+vp9/decoder/vp9_reader.c
+vp9/decoder/vp9_reader.h
+vp9/decoder/vp9_thread.c
+vp9/decoder/vp9_thread.h
+vp9/vp9_common.mk
+vp9/vp9_dx_iface.c
+vp9/vp9dx.mk
+vp9/vp9_iface_common.h
+vpx_config.c
+vpx/internal/vpx_codec_internal.h
+vpx/internal/vpx_psnr.h
+vpx_mem/include/vpx_mem_intrnl.h
+vpx_mem/vpx_mem.c
+vpx_mem/vpx_mem.h
+vpx_mem/vpx_mem.mk
+vpx_ports/asm_offsets.h
+vpx_ports/emmintrin_compat.h
+vpx_ports/emms.asm
+vpx_ports/mem.h
+vpx_ports/mem_ops_aligned.h
+vpx_ports/mem_ops.h
+vpx_ports/vpx_once.h
+vpx_ports/vpx_ports.mk
+vpx_ports/vpx_timer.h
+vpx_ports/x86_abi_support.asm
+vpx_ports/x86.h
+vpx_scale/generic/gen_scalers.c
+vpx_scale/generic/vpx_scale.c
+vpx_scale/generic/yv12config.c
+vpx_scale/generic/yv12extend.c
+vpx_scale/vpx_scale_asm_offsets.c
+vpx_scale/vpx_scale.h
+vpx_scale/vpx_scale.mk
+vpx_scale/vpx_scale_rtcd.c
+vpx_scale/vpx_scale_rtcd.pl
+vpx_scale/yv12config.h
+vpx/src/vpx_codec.c
+vpx/src/vpx_decoder.c
+vpx/src/vpx_encoder.c
+vpx/src/vpx_image.c
+vpx/src/vpx_psnr.c
+vpx/vp8cx.h
+vpx/vp8dx.h
+vpx/vp8.h
+vpx/vpx_codec.h
+vpx/vpx_codec.mk
+vpx/vpx_decoder.h
+vpx/vpx_encoder.h
+vpx/vpx_frame_buffer.h
+vpx/vpx_image.h
+vpx/vpx_integer.h
diff --git a/x86/vp8_rtcd.h b/x86/vp8_rtcd.h
new file mode 100644
index 0000000..720aa3c
--- /dev/null
+++ b/x86/vp8_rtcd.h
@@ -0,0 +1,468 @@
+#ifndef VP8_RTCD_H_
+#define VP8_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*
+ * VP8
+ */
+
+struct blockd;
+struct macroblockd;
+struct loop_filter_info;
+
+/* Encoder forward decls */
+struct block;
+struct macroblock;
+struct variance_vtable;
+union int_mv;
+struct yv12_buffer_config;
+
+void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict16x16_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict16x16_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict16x16 vp8_bilinear_predict16x16_sse2
+
+void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict4x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_mmx
+
+void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_mmx
+
+void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x8_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_bilinear_predict8x8_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_bilinear_predict8x8 vp8_bilinear_predict8x8_sse2
+
+void vp8_blend_b_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
+#define vp8_blend_b vp8_blend_b_c
+
+void vp8_blend_mb_inner_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
+#define vp8_blend_mb_inner vp8_blend_mb_inner_c
+
+void vp8_blend_mb_outer_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
+#define vp8_blend_mb_outer vp8_blend_mb_outer_c
+
+int vp8_block_error_c(short *coeff, short *dqcoeff);
+int vp8_block_error_mmx(short *coeff, short *dqcoeff);
+int vp8_block_error_xmm(short *coeff, short *dqcoeff);
+#define vp8_block_error vp8_block_error_xmm
+
+void vp8_build_intra_predictors_mbuv_s_c(struct macroblockd *x, unsigned char * uabove_row, unsigned char * vabove_row,  unsigned char *uleft, unsigned char *vleft, int left_stride, unsigned char * upred_ptr, unsigned char * vpred_ptr, int pred_stride);
+void vp8_build_intra_predictors_mbuv_s_sse2(struct macroblockd *x, unsigned char * uabove_row, unsigned char * vabove_row,  unsigned char *uleft, unsigned char *vleft, int left_stride, unsigned char * upred_ptr, unsigned char * vpred_ptr, int pred_stride);
+#define vp8_build_intra_predictors_mbuv_s vp8_build_intra_predictors_mbuv_s_sse2
+
+void vp8_build_intra_predictors_mby_s_c(struct macroblockd *x, unsigned char * yabove_row, unsigned char * yleft, int left_stride, unsigned char * ypred_ptr, int y_stride);
+void vp8_build_intra_predictors_mby_s_sse2(struct macroblockd *x, unsigned char * yabove_row, unsigned char * yleft, int left_stride, unsigned char * ypred_ptr, int y_stride);
+#define vp8_build_intra_predictors_mby_s vp8_build_intra_predictors_mby_s_sse2
+
+void vp8_clear_system_state_c();
+void vpx_reset_mmx_state();
+#define vp8_clear_system_state vpx_reset_mmx_state
+
+void vp8_copy32xn_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, int n);
+void vp8_copy32xn_sse2(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, int n);
+#define vp8_copy32xn vp8_copy32xn_sse2
+
+void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem16x16_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem16x16_sse2(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem16x16 vp8_copy_mem16x16_sse2
+
+void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x4_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x4 vp8_copy_mem8x4_mmx
+
+void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp8_copy_mem8x8_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+#define vp8_copy_mem8x8 vp8_copy_mem8x8_mmx
+
+void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+void vp8_dc_only_idct_add_mmx(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+#define vp8_dc_only_idct_add vp8_dc_only_idct_add_mmx
+
+int vp8_denoiser_filter_c(struct yv12_buffer_config* mc_running_avg, struct yv12_buffer_config* running_avg, struct macroblock* signal, unsigned int motion_magnitude2, int y_offset, int uv_offset);
+int vp8_denoiser_filter_sse2(struct yv12_buffer_config* mc_running_avg, struct yv12_buffer_config* running_avg, struct macroblock* signal, unsigned int motion_magnitude2, int y_offset, int uv_offset);
+#define vp8_denoiser_filter vp8_denoiser_filter_sse2
+
+void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
+void vp8_dequant_idct_add_mmx(short *input, short *dq, unsigned char *output, int stride);
+#define vp8_dequant_idct_add vp8_dequant_idct_add_mmx
+
+void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+void vp8_dequant_idct_add_uv_block_mmx(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+void vp8_dequant_idct_add_uv_block_sse2(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+#define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_sse2
+
+void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+void vp8_dequant_idct_add_y_block_mmx(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+void vp8_dequant_idct_add_y_block_sse2(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+#define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_sse2
+
+void vp8_dequantize_b_c(struct blockd*, short *dqc);
+void vp8_dequantize_b_mmx(struct blockd*, short *dqc);
+#define vp8_dequantize_b vp8_dequantize_b_mmx
+
+int vp8_diamond_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_diamond_search_sad vp8_diamond_search_sad_c
+
+void vp8_fast_quantize_b_c(struct block *, struct blockd *);
+void vp8_fast_quantize_b_sse2(struct block *, struct blockd *);
+#define vp8_fast_quantize_b vp8_fast_quantize_b_sse2
+
+void vp8_fast_quantize_b_pair_c(struct block *b1, struct block *b2, struct blockd *d1, struct blockd *d2);
+#define vp8_fast_quantize_b_pair vp8_fast_quantize_b_pair_c
+
+void vp8_filter_by_weight16x16_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+void vp8_filter_by_weight16x16_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+#define vp8_filter_by_weight16x16 vp8_filter_by_weight16x16_sse2
+
+void vp8_filter_by_weight4x4_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+#define vp8_filter_by_weight4x4 vp8_filter_by_weight4x4_c
+
+void vp8_filter_by_weight8x8_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+void vp8_filter_by_weight8x8_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+#define vp8_filter_by_weight8x8 vp8_filter_by_weight8x8_sse2
+
+int vp8_full_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_full_search_sad vp8_full_search_sad_c
+
+unsigned int vp8_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+unsigned int vp8_get4x4sse_cs_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+#define vp8_get4x4sse_cs vp8_get4x4sse_cs_mmx
+
+unsigned int vp8_get_mb_ss_c(const short *);
+unsigned int vp8_get_mb_ss_mmx(const short *);
+unsigned int vp8_get_mb_ss_sse2(const short *);
+#define vp8_get_mb_ss vp8_get_mb_ss_sse2
+
+void vp8_intra4x4_predict_c(unsigned char *Above, unsigned char *yleft, int left_stride, int b_mode, unsigned char *dst, int dst_stride, unsigned char top_left);
+#define vp8_intra4x4_predict vp8_intra4x4_predict_c
+
+void vp8_loop_filter_bh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bh_mmx(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bh_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bh vp8_loop_filter_bh_sse2
+
+void vp8_loop_filter_bv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bv_mmx(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_bv_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_bv vp8_loop_filter_bv_sse2
+
+void vp8_loop_filter_mbh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbh_mmx(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbh_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbh vp8_loop_filter_mbh_sse2
+
+void vp8_loop_filter_mbv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbv_mmx(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp8_loop_filter_mbv_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+#define vp8_loop_filter_mbv vp8_loop_filter_mbv_sse2
+
+void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bhs_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bhs_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_sse2
+
+void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bvs_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_bvs_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_sse2
+
+void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_horizontal_edge_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_horizontal_edge_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbh vp8_loop_filter_simple_horizontal_edge_sse2
+
+void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_vertical_edge_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
+void vp8_loop_filter_simple_vertical_edge_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
+#define vp8_loop_filter_simple_mbv vp8_loop_filter_simple_vertical_edge_sse2
+
+int vp8_mbblock_error_c(struct macroblock *mb, int dc);
+int vp8_mbblock_error_mmx(struct macroblock *mb, int dc);
+int vp8_mbblock_error_xmm(struct macroblock *mb, int dc);
+#define vp8_mbblock_error vp8_mbblock_error_xmm
+
+void vp8_mbpost_proc_across_ip_c(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+void vp8_mbpost_proc_across_ip_xmm(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+#define vp8_mbpost_proc_across_ip vp8_mbpost_proc_across_ip_xmm
+
+void vp8_mbpost_proc_down_c(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+void vp8_mbpost_proc_down_mmx(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+void vp8_mbpost_proc_down_xmm(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+#define vp8_mbpost_proc_down vp8_mbpost_proc_down_xmm
+
+int vp8_mbuverror_c(struct macroblock *mb);
+int vp8_mbuverror_mmx(struct macroblock *mb);
+int vp8_mbuverror_xmm(struct macroblock *mb);
+#define vp8_mbuverror vp8_mbuverror_xmm
+
+unsigned int vp8_mse16x16_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_mse16x16_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_mse16x16_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+#define vp8_mse16x16 vp8_mse16x16_wmt
+
+void vp8_plane_add_noise_c(unsigned char *s, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int w, unsigned int h, int pitch);
+void vp8_plane_add_noise_mmx(unsigned char *s, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int w, unsigned int h, int pitch);
+void vp8_plane_add_noise_wmt(unsigned char *s, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int w, unsigned int h, int pitch);
+#define vp8_plane_add_noise vp8_plane_add_noise_wmt
+
+void vp8_post_proc_down_and_across_mb_row_c(unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size);
+void vp8_post_proc_down_and_across_mb_row_sse2(unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size);
+#define vp8_post_proc_down_and_across_mb_row vp8_post_proc_down_and_across_mb_row_sse2
+
+void vp8_quantize_mb_c(struct macroblock *);
+#define vp8_quantize_mb vp8_quantize_mb_c
+
+void vp8_quantize_mbuv_c(struct macroblock *);
+#define vp8_quantize_mbuv vp8_quantize_mbuv_c
+
+void vp8_quantize_mby_c(struct macroblock *);
+#define vp8_quantize_mby vp8_quantize_mby_c
+
+int vp8_refining_search_sad_c(struct macroblock *x, struct block *b, struct blockd *d, union int_mv *ref_mv, int sad_per_bit, int distance, struct variance_vtable *fn_ptr, int *mvcost[2], union int_mv *center_mv);
+#define vp8_refining_search_sad vp8_refining_search_sad_c
+
+void vp8_regular_quantize_b_c(struct block *, struct blockd *);
+void vp8_regular_quantize_b_sse2(struct block *, struct blockd *);
+#define vp8_regular_quantize_b vp8_regular_quantize_b_sse2
+
+void vp8_regular_quantize_b_pair_c(struct block *b1, struct block *b2, struct blockd *d1, struct blockd *d2);
+#define vp8_regular_quantize_b_pair vp8_regular_quantize_b_pair_c
+
+unsigned int vp8_sad16x16_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp8_sad16x16_mmx(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp8_sad16x16_wmt(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+#define vp8_sad16x16 vp8_sad16x16_wmt
+
+void vp8_sad16x16x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp8_sad16x16x3 vp8_sad16x16x3_c
+
+void vp8_sad16x16x4d_c(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp8_sad16x16x4d vp8_sad16x16x4d_c
+
+void vp8_sad16x16x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+#define vp8_sad16x16x8 vp8_sad16x16x8_c
+
+unsigned int vp8_sad16x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp8_sad16x8_mmx(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp8_sad16x8_wmt(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+#define vp8_sad16x8 vp8_sad16x8_wmt
+
+void vp8_sad16x8x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp8_sad16x8x3 vp8_sad16x8x3_c
+
+void vp8_sad16x8x4d_c(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp8_sad16x8x4d vp8_sad16x8x4d_c
+
+void vp8_sad16x8x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+#define vp8_sad16x8x8 vp8_sad16x8x8_c
+
+unsigned int vp8_sad4x4_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp8_sad4x4_mmx(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp8_sad4x4_wmt(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+#define vp8_sad4x4 vp8_sad4x4_wmt
+
+void vp8_sad4x4x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp8_sad4x4x3 vp8_sad4x4x3_c
+
+void vp8_sad4x4x4d_c(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp8_sad4x4x4d vp8_sad4x4x4d_c
+
+void vp8_sad4x4x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+#define vp8_sad4x4x8 vp8_sad4x4x8_c
+
+unsigned int vp8_sad8x16_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp8_sad8x16_mmx(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp8_sad8x16_wmt(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+#define vp8_sad8x16 vp8_sad8x16_wmt
+
+void vp8_sad8x16x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp8_sad8x16x3 vp8_sad8x16x3_c
+
+void vp8_sad8x16x4d_c(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp8_sad8x16x4d vp8_sad8x16x4d_c
+
+void vp8_sad8x16x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+#define vp8_sad8x16x8 vp8_sad8x16x8_c
+
+unsigned int vp8_sad8x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp8_sad8x8_mmx(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp8_sad8x8_wmt(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+#define vp8_sad8x8 vp8_sad8x8_wmt
+
+void vp8_sad8x8x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp8_sad8x8x3 vp8_sad8x8x3_c
+
+void vp8_sad8x8x4d_c(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp8_sad8x8x4d vp8_sad8x8x4d_c
+
+void vp8_sad8x8x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+#define vp8_sad8x8x8 vp8_sad8x8x8_c
+
+void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct4x4_mmx(short *input, short *output, int pitch);
+void vp8_short_fdct4x4_sse2(short *input, short *output, int pitch);
+#define vp8_short_fdct4x4 vp8_short_fdct4x4_sse2
+
+void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
+void vp8_short_fdct8x4_mmx(short *input, short *output, int pitch);
+void vp8_short_fdct8x4_sse2(short *input, short *output, int pitch);
+#define vp8_short_fdct8x4 vp8_short_fdct8x4_sse2
+
+void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+void vp8_short_idct4x4llm_mmx(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+#define vp8_short_idct4x4llm vp8_short_idct4x4llm_mmx
+
+void vp8_short_inv_walsh4x4_c(short *input, short *output);
+void vp8_short_inv_walsh4x4_mmx(short *input, short *output);
+void vp8_short_inv_walsh4x4_sse2(short *input, short *output);
+#define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_sse2
+
+void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
+
+void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
+void vp8_short_walsh4x4_sse2(short *input, short *output, int pitch);
+#define vp8_short_walsh4x4 vp8_short_walsh4x4_sse2
+
+void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict16x16_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict16x16 vp8_sixtap_predict16x16_sse2
+
+void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict4x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_mmx
+
+void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x4_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x4 vp8_sixtap_predict8x4_sse2
+
+void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+void vp8_sixtap_predict8x8_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+#define vp8_sixtap_predict8x8 vp8_sixtap_predict8x8_sse2
+
+unsigned int vp8_sub_pixel_mse16x16_c(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+unsigned int vp8_sub_pixel_mse16x16_mmx(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+unsigned int vp8_sub_pixel_mse16x16_wmt(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+#define vp8_sub_pixel_mse16x16 vp8_sub_pixel_mse16x16_wmt
+
+unsigned int vp8_sub_pixel_variance16x16_c(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+unsigned int vp8_sub_pixel_variance16x16_mmx(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+unsigned int vp8_sub_pixel_variance16x16_wmt(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+#define vp8_sub_pixel_variance16x16 vp8_sub_pixel_variance16x16_wmt
+
+unsigned int vp8_sub_pixel_variance16x8_c(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+unsigned int vp8_sub_pixel_variance16x8_mmx(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+unsigned int vp8_sub_pixel_variance16x8_wmt(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+#define vp8_sub_pixel_variance16x8 vp8_sub_pixel_variance16x8_wmt
+
+unsigned int vp8_sub_pixel_variance4x4_c(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+unsigned int vp8_sub_pixel_variance4x4_mmx(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+unsigned int vp8_sub_pixel_variance4x4_wmt(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+#define vp8_sub_pixel_variance4x4 vp8_sub_pixel_variance4x4_wmt
+
+unsigned int vp8_sub_pixel_variance8x16_c(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+unsigned int vp8_sub_pixel_variance8x16_mmx(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+unsigned int vp8_sub_pixel_variance8x16_wmt(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+#define vp8_sub_pixel_variance8x16 vp8_sub_pixel_variance8x16_wmt
+
+unsigned int vp8_sub_pixel_variance8x8_c(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+unsigned int vp8_sub_pixel_variance8x8_mmx(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+unsigned int vp8_sub_pixel_variance8x8_wmt(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+#define vp8_sub_pixel_variance8x8 vp8_sub_pixel_variance8x8_wmt
+
+void vp8_subtract_b_c(struct block *be, struct blockd *bd, int pitch);
+void vp8_subtract_b_mmx(struct block *be, struct blockd *bd, int pitch);
+void vp8_subtract_b_sse2(struct block *be, struct blockd *bd, int pitch);
+#define vp8_subtract_b vp8_subtract_b_sse2
+
+void vp8_subtract_mbuv_c(short *diff, unsigned char *usrc, unsigned char *vsrc, int src_stride, unsigned char *upred, unsigned char *vpred, int pred_stride);
+void vp8_subtract_mbuv_mmx(short *diff, unsigned char *usrc, unsigned char *vsrc, int src_stride, unsigned char *upred, unsigned char *vpred, int pred_stride);
+void vp8_subtract_mbuv_sse2(short *diff, unsigned char *usrc, unsigned char *vsrc, int src_stride, unsigned char *upred, unsigned char *vpred, int pred_stride);
+#define vp8_subtract_mbuv vp8_subtract_mbuv_sse2
+
+void vp8_subtract_mby_c(short *diff, unsigned char *src, int src_stride, unsigned char *pred, int pred_stride);
+void vp8_subtract_mby_mmx(short *diff, unsigned char *src, int src_stride, unsigned char *pred, int pred_stride);
+void vp8_subtract_mby_sse2(short *diff, unsigned char *src, int src_stride, unsigned char *pred, int pred_stride);
+#define vp8_subtract_mby vp8_subtract_mby_sse2
+
+void vp8_temporal_filter_apply_c(unsigned char *frame1, unsigned int stride, unsigned char *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, unsigned short *count);
+void vp8_temporal_filter_apply_sse2(unsigned char *frame1, unsigned int stride, unsigned char *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, unsigned short *count);
+#define vp8_temporal_filter_apply vp8_temporal_filter_apply_sse2
+
+unsigned int vp8_variance16x16_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance16x16_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance16x16_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+#define vp8_variance16x16 vp8_variance16x16_wmt
+
+unsigned int vp8_variance16x8_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance16x8_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance16x8_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+#define vp8_variance16x8 vp8_variance16x8_wmt
+
+unsigned int vp8_variance4x4_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance4x4_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance4x4_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+#define vp8_variance4x4 vp8_variance4x4_wmt
+
+unsigned int vp8_variance8x16_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance8x16_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance8x16_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+#define vp8_variance8x16 vp8_variance8x16_wmt
+
+unsigned int vp8_variance8x8_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance8x8_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance8x8_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+#define vp8_variance8x8 vp8_variance8x8_wmt
+
+unsigned int vp8_variance_halfpixvar16x16_h_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance_halfpixvar16x16_h_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance_halfpixvar16x16_h_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+#define vp8_variance_halfpixvar16x16_h vp8_variance_halfpixvar16x16_h_wmt
+
+unsigned int vp8_variance_halfpixvar16x16_hv_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance_halfpixvar16x16_hv_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance_halfpixvar16x16_hv_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+#define vp8_variance_halfpixvar16x16_hv vp8_variance_halfpixvar16x16_hv_wmt
+
+unsigned int vp8_variance_halfpixvar16x16_v_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance_halfpixvar16x16_v_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+unsigned int vp8_variance_halfpixvar16x16_v_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+#define vp8_variance_halfpixvar16x16_v vp8_variance_halfpixvar16x16_v_wmt
+
+void vp8_yv12_copy_partial_frame_c(struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vp8_yv12_copy_partial_frame vp8_yv12_copy_partial_frame_c
+
+void vp8_rtcd(void);
+
+#ifdef RTCD_C
+#include "vpx_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/x86/vp9_rtcd.h b/x86/vp9_rtcd.h
new file mode 100644
index 0000000..3f0a307
--- /dev/null
+++ b/x86/vp9_rtcd.h
@@ -0,0 +1,363 @@
+#ifndef VP9_RTCD_H_
+#define VP9_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*
+ * VP9
+ */
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_enums.h"
+
+struct macroblockd;
+
+/* Encoder forward decls */
+struct macroblock;
+struct vp9_variance_vtable;
+
+#define DEC_MVCOSTS int *mvjcost, int *mvcost[2]
+struct mv;
+union int_mv;
+struct yv12_buffer_config;
+
+void vp9_blend_b_c(uint8_t *y, uint8_t *u, uint8_t *v, int y1, int u1, int v1, int alpha, int stride);
+#define vp9_blend_b vp9_blend_b_c
+
+void vp9_blend_mb_inner_c(uint8_t *y, uint8_t *u, uint8_t *v, int y1, int u1, int v1, int alpha, int stride);
+#define vp9_blend_mb_inner vp9_blend_mb_inner_c
+
+void vp9_blend_mb_outer_c(uint8_t *y, uint8_t *u, uint8_t *v, int y1, int u1, int v1, int alpha, int stride);
+#define vp9_blend_mb_outer vp9_blend_mb_outer_c
+
+void vp9_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 vp9_convolve8_sse2(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 vp9_convolve8 vp9_convolve8_sse2
+
+void vp9_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 vp9_convolve8_avg_sse2(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 vp9_convolve8_avg vp9_convolve8_avg_sse2
+
+void vp9_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 vp9_convolve8_avg_horiz_sse2(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 vp9_convolve8_avg_horiz vp9_convolve8_avg_horiz_sse2
+
+void vp9_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 vp9_convolve8_avg_vert_sse2(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 vp9_convolve8_avg_vert vp9_convolve8_avg_vert_sse2
+
+void vp9_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 vp9_convolve8_horiz_sse2(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 vp9_convolve8_horiz vp9_convolve8_horiz_sse2
+
+void vp9_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 vp9_convolve8_vert_sse2(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 vp9_convolve8_vert vp9_convolve8_vert_sse2
+
+void vp9_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 vp9_convolve_avg_sse2(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 vp9_convolve_avg vp9_convolve_avg_sse2
+
+void vp9_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 vp9_convolve_copy_sse2(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 vp9_convolve_copy vp9_convolve_copy_sse2
+
+void vp9_d117_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d117_predictor_16x16 vp9_d117_predictor_16x16_c
+
+void vp9_d117_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d117_predictor_32x32 vp9_d117_predictor_32x32_c
+
+void vp9_d117_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d117_predictor_4x4 vp9_d117_predictor_4x4_c
+
+void vp9_d117_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d117_predictor_8x8 vp9_d117_predictor_8x8_c
+
+void vp9_d135_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d135_predictor_16x16 vp9_d135_predictor_16x16_c
+
+void vp9_d135_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d135_predictor_32x32 vp9_d135_predictor_32x32_c
+
+void vp9_d135_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d135_predictor_4x4 vp9_d135_predictor_4x4_c
+
+void vp9_d135_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d135_predictor_8x8 vp9_d135_predictor_8x8_c
+
+void vp9_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d153_predictor_16x16 vp9_d153_predictor_16x16_c
+
+void vp9_d153_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d153_predictor_32x32 vp9_d153_predictor_32x32_c
+
+void vp9_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d153_predictor_4x4 vp9_d153_predictor_4x4_c
+
+void vp9_d153_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d153_predictor_8x8 vp9_d153_predictor_8x8_c
+
+void vp9_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d207_predictor_16x16 vp9_d207_predictor_16x16_c
+
+void vp9_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d207_predictor_32x32 vp9_d207_predictor_32x32_c
+
+void vp9_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d207_predictor_4x4 vp9_d207_predictor_4x4_c
+
+void vp9_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d207_predictor_8x8 vp9_d207_predictor_8x8_c
+
+void vp9_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d45_predictor_16x16 vp9_d45_predictor_16x16_c
+
+void vp9_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d45_predictor_32x32 vp9_d45_predictor_32x32_c
+
+void vp9_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d45_predictor_4x4 vp9_d45_predictor_4x4_c
+
+void vp9_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d45_predictor_8x8 vp9_d45_predictor_8x8_c
+
+void vp9_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d63_predictor_16x16 vp9_d63_predictor_16x16_c
+
+void vp9_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d63_predictor_32x32 vp9_d63_predictor_32x32_c
+
+void vp9_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d63_predictor_4x4 vp9_d63_predictor_4x4_c
+
+void vp9_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_d63_predictor_8x8 vp9_d63_predictor_8x8_c
+
+void vp9_dc_128_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_128_predictor_16x16 vp9_dc_128_predictor_16x16_c
+
+void vp9_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_128_predictor_32x32 vp9_dc_128_predictor_32x32_c
+
+void vp9_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_128_predictor_4x4 vp9_dc_128_predictor_4x4_c
+
+void vp9_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_128_predictor_8x8 vp9_dc_128_predictor_8x8_c
+
+void vp9_dc_left_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_left_predictor_16x16 vp9_dc_left_predictor_16x16_c
+
+void vp9_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_left_predictor_32x32 vp9_dc_left_predictor_32x32_c
+
+void vp9_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_left_predictor_4x4 vp9_dc_left_predictor_4x4_c
+
+void vp9_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_left_predictor_8x8 vp9_dc_left_predictor_8x8_c
+
+void vp9_dc_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_dc_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_predictor_16x16 vp9_dc_predictor_16x16_sse2
+
+void vp9_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_dc_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_predictor_32x32 vp9_dc_predictor_32x32_sse2
+
+void vp9_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_dc_predictor_4x4_sse(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_predictor_4x4 vp9_dc_predictor_4x4_sse
+
+void vp9_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_dc_predictor_8x8_sse(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_predictor_8x8 vp9_dc_predictor_8x8_sse
+
+void vp9_dc_top_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_top_predictor_16x16 vp9_dc_top_predictor_16x16_c
+
+void vp9_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_top_predictor_32x32 vp9_dc_top_predictor_32x32_c
+
+void vp9_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_top_predictor_4x4 vp9_dc_top_predictor_4x4_c
+
+void vp9_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_dc_top_predictor_8x8 vp9_dc_top_predictor_8x8_c
+
+void vp9_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_h_predictor_16x16 vp9_h_predictor_16x16_c
+
+void vp9_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_h_predictor_32x32 vp9_h_predictor_32x32_c
+
+void vp9_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_h_predictor_4x4 vp9_h_predictor_4x4_c
+
+void vp9_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_h_predictor_8x8 vp9_h_predictor_8x8_c
+
+void vp9_idct16x16_10_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+void vp9_idct16x16_10_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct16x16_10_add vp9_idct16x16_10_add_sse2
+
+void vp9_idct16x16_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+void vp9_idct16x16_1_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct16x16_1_add vp9_idct16x16_1_add_sse2
+
+void vp9_idct16x16_256_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+void vp9_idct16x16_256_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct16x16_256_add vp9_idct16x16_256_add_sse2
+
+void vp9_idct32x32_1024_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+void vp9_idct32x32_1024_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct32x32_1024_add vp9_idct32x32_1024_add_sse2
+
+void vp9_idct32x32_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+void vp9_idct32x32_1_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct32x32_1_add vp9_idct32x32_1_add_sse2
+
+void vp9_idct32x32_34_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+void vp9_idct32x32_34_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct32x32_34_add vp9_idct32x32_34_add_sse2
+
+void vp9_idct4x4_16_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+void vp9_idct4x4_16_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct4x4_16_add vp9_idct4x4_16_add_sse2
+
+void vp9_idct4x4_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+void vp9_idct4x4_1_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct4x4_1_add vp9_idct4x4_1_add_sse2
+
+void vp9_idct8x8_10_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+void vp9_idct8x8_10_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct8x8_10_add vp9_idct8x8_10_add_sse2
+
+void vp9_idct8x8_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+void vp9_idct8x8_1_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct8x8_1_add vp9_idct8x8_1_add_sse2
+
+void vp9_idct8x8_64_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+void vp9_idct8x8_64_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_idct8x8_64_add vp9_idct8x8_64_add_sse2
+
+void vp9_iht16x16_256_add_c(const int16_t *input, uint8_t *output, int pitch, int tx_type);
+void vp9_iht16x16_256_add_sse2(const int16_t *input, uint8_t *output, int pitch, int tx_type);
+#define vp9_iht16x16_256_add vp9_iht16x16_256_add_sse2
+
+void vp9_iht4x4_16_add_c(const int16_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht4x4_16_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht4x4_16_add vp9_iht4x4_16_add_sse2
+
+void vp9_iht8x8_64_add_c(const int16_t *input, uint8_t *dest, int dest_stride, int tx_type);
+void vp9_iht8x8_64_add_sse2(const int16_t *input, uint8_t *dest, int dest_stride, int tx_type);
+#define vp9_iht8x8_64_add vp9_iht8x8_64_add_sse2
+
+void vp9_iwht4x4_16_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_iwht4x4_16_add vp9_iwht4x4_16_add_c
+
+void vp9_iwht4x4_1_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_iwht4x4_1_add vp9_iwht4x4_1_add_c
+
+void vp9_lpf_horizontal_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_lpf_horizontal_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+#define vp9_lpf_horizontal_16 vp9_lpf_horizontal_16_sse2
+
+void vp9_lpf_horizontal_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_lpf_horizontal_4_mmx(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+#define vp9_lpf_horizontal_4 vp9_lpf_horizontal_4_mmx
+
+void vp9_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vp9_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vp9_lpf_horizontal_4_dual vp9_lpf_horizontal_4_dual_sse2
+
+void vp9_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+#define vp9_lpf_horizontal_8 vp9_lpf_horizontal_8_sse2
+
+void vp9_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vp9_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vp9_lpf_horizontal_8_dual vp9_lpf_horizontal_8_dual_sse2
+
+void vp9_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vp9_lpf_vertical_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vp9_lpf_vertical_16 vp9_lpf_vertical_16_sse2
+
+void vp9_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vp9_lpf_vertical_16_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vp9_lpf_vertical_16_dual vp9_lpf_vertical_16_dual_sse2
+
+void vp9_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_lpf_vertical_4_mmx(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+#define vp9_lpf_vertical_4 vp9_lpf_vertical_4_mmx
+
+void vp9_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vp9_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vp9_lpf_vertical_4_dual vp9_lpf_vertical_4_dual_sse2
+
+void vp9_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+#define vp9_lpf_vertical_8 vp9_lpf_vertical_8_sse2
+
+void vp9_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+void vp9_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
+#define vp9_lpf_vertical_8_dual vp9_lpf_vertical_8_dual_sse2
+
+void vp9_tm_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_tm_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_tm_predictor_16x16 vp9_tm_predictor_16x16_sse2
+
+void vp9_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_tm_predictor_32x32 vp9_tm_predictor_32x32_c
+
+void vp9_tm_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_tm_predictor_4x4_sse(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_tm_predictor_4x4 vp9_tm_predictor_4x4_sse
+
+void vp9_tm_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_tm_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_tm_predictor_8x8 vp9_tm_predictor_8x8_sse2
+
+void vp9_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_v_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_v_predictor_16x16 vp9_v_predictor_16x16_sse2
+
+void vp9_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_v_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_v_predictor_32x32 vp9_v_predictor_32x32_sse2
+
+void vp9_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_v_predictor_4x4_sse(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_v_predictor_4x4 vp9_v_predictor_4x4_sse
+
+void vp9_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_v_predictor_8x8_sse(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_v_predictor_8x8 vp9_v_predictor_8x8_sse
+
+void vp9_rtcd(void);
+
+#ifdef RTCD_C
+#include "vpx_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/x86/vpx_config.asm b/x86/vpx_config.asm
new file mode 100644
index 0000000..4b3bb2a
--- /dev/null
+++ b/x86/vpx_config.asm
@@ -0,0 +1,80 @@
+ARCH_ARM equ 0
+ARCH_MIPS equ 0
+ARCH_X86 equ 1
+ARCH_X86_64 equ 0
+ARCH_PPC32 equ 0
+ARCH_PPC64 equ 0
+HAVE_EDSP equ 0
+HAVE_MEDIA equ 0
+HAVE_NEON equ 0
+HAVE_MIPS32 equ 0
+HAVE_DSPR2 equ 0
+HAVE_MMX equ 1
+HAVE_SSE equ 1
+HAVE_SSE2 equ 1
+HAVE_SSE3 equ 0
+HAVE_SSSE3 equ 0
+HAVE_SSE4_1 equ 0
+HAVE_AVX equ 0
+HAVE_AVX2 equ 0
+HAVE_ALTIVEC equ 0
+HAVE_VPX_PORTS equ 1
+HAVE_STDINT_H equ 1
+HAVE_ALT_TREE_LAYOUT equ 0
+HAVE_PTHREAD_H equ 1
+HAVE_SYS_MMAN_H equ 1
+HAVE_UNISTD_H equ 1
+CONFIG_EXTERNAL_BUILD equ 0
+CONFIG_INSTALL_DOCS equ 1
+CONFIG_INSTALL_BINS equ 1
+CONFIG_INSTALL_LIBS equ 1
+CONFIG_INSTALL_SRCS equ 0
+CONFIG_USE_X86INC equ 1
+CONFIG_DEBUG equ 0
+CONFIG_GPROF equ 0
+CONFIG_GCOV equ 0
+CONFIG_RVCT equ 0
+CONFIG_GCC equ 1
+CONFIG_MSVS equ 0
+CONFIG_PIC equ 1
+CONFIG_BIG_ENDIAN equ 0
+CONFIG_CODEC_SRCS equ 0
+CONFIG_DEBUG_LIBS equ 0
+CONFIG_FAST_UNALIGNED equ 1
+CONFIG_MEM_MANAGER equ 0
+CONFIG_MEM_TRACKER equ 0
+CONFIG_MEM_CHECKS equ 0
+CONFIG_DEQUANT_TOKENS equ 0
+CONFIG_DC_RECON equ 0
+CONFIG_RUNTIME_CPU_DETECT equ 0
+CONFIG_POSTPROC equ 1
+CONFIG_VP9_POSTPROC equ 0
+CONFIG_MULTITHREAD equ 1
+CONFIG_INTERNAL_STATS equ 0
+CONFIG_VP8_ENCODER equ 1
+CONFIG_VP8_DECODER equ 1
+CONFIG_VP9_ENCODER equ 0
+CONFIG_VP9_DECODER equ 1
+CONFIG_VP8 equ 1
+CONFIG_VP9 equ 1
+CONFIG_ENCODERS equ 1
+CONFIG_DECODERS equ 1
+CONFIG_STATIC_MSVCRT equ 0
+CONFIG_SPATIAL_RESAMPLING equ 1
+CONFIG_REALTIME_ONLY equ 0
+CONFIG_ONTHEFLY_BITPACKING equ 0
+CONFIG_ERROR_CONCEALMENT equ 0
+CONFIG_SHARED equ 0
+CONFIG_STATIC equ 1
+CONFIG_SMALL equ 0
+CONFIG_POSTPROC_VISUALIZER equ 0
+CONFIG_OS_SUPPORT equ 1
+CONFIG_UNIT_TESTS equ 0
+CONFIG_WEBM_IO equ 1
+CONFIG_DECODE_PERF_TESTS equ 0
+CONFIG_MULTI_RES_ENCODING equ 0
+CONFIG_TEMPORAL_DENOISING equ 1
+CONFIG_EXPERIMENTAL equ 0
+CONFIG_DECRYPT equ 0
+CONFIG_MULTIPLE_ARF equ 0
+CONFIG_ALPHA equ 0
diff --git a/x86/vpx_config.c b/x86/vpx_config.c
new file mode 100644
index 0000000..3a9faed
--- /dev/null
+++ b/x86/vpx_config.c
@@ -0,0 +1,9 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+static const char* const cfg = "--force-target=x86-android-gcc --disable-runtime-cpu-detect --sdk-path=/usr/local/google/home/johannkoenig/android-ndk --disable-vp9-encoder --disable-examples --disable-docs --disable-sse3 --disable-ssse3 --disable-sse4_1 --disable-avx --disable-avx2 --enable-pic";
+const char *vpx_codec_build_config(void) {return cfg;}
diff --git a/x86/vpx_config.h b/x86/vpx_config.h
new file mode 100644
index 0000000..be04367
--- /dev/null
+++ b/x86/vpx_config.h
@@ -0,0 +1,93 @@
+/* Copyright (c) 2011 The WebM project authors. All Rights Reserved. */
+/*  */
+/* Use of this source code is governed by a BSD-style license */
+/* that can be found in the LICENSE file in the root of the source */
+/* tree. An additional intellectual property rights grant can be found */
+/* in the file PATENTS.  All contributing project authors may */
+/* be found in the AUTHORS file in the root of the source tree. */
+/* This file automatically generated by configure. Do not edit! */
+#ifndef VPX_CONFIG_H
+#define VPX_CONFIG_H
+#define RESTRICT    
+#define INLINE      __inline__ __attribute__((always_inline))
+#define ARCH_ARM 0
+#define ARCH_MIPS 0
+#define ARCH_X86 1
+#define ARCH_X86_64 0
+#define ARCH_PPC32 0
+#define ARCH_PPC64 0
+#define HAVE_EDSP 0
+#define HAVE_MEDIA 0
+#define HAVE_NEON 0
+#define HAVE_MIPS32 0
+#define HAVE_DSPR2 0
+#define HAVE_MMX 1
+#define HAVE_SSE 1
+#define HAVE_SSE2 1
+#define HAVE_SSE3 0
+#define HAVE_SSSE3 0
+#define HAVE_SSE4_1 0
+#define HAVE_AVX 0
+#define HAVE_AVX2 0
+#define HAVE_ALTIVEC 0
+#define HAVE_VPX_PORTS 1
+#define HAVE_STDINT_H 1
+#define HAVE_ALT_TREE_LAYOUT 0
+#define HAVE_PTHREAD_H 1
+#define HAVE_SYS_MMAN_H 1
+#define HAVE_UNISTD_H 1
+#define CONFIG_EXTERNAL_BUILD 0
+#define CONFIG_INSTALL_DOCS 1
+#define CONFIG_INSTALL_BINS 1
+#define CONFIG_INSTALL_LIBS 1
+#define CONFIG_INSTALL_SRCS 0
+#define CONFIG_USE_X86INC 1
+#define CONFIG_DEBUG 0
+#define CONFIG_GPROF 0
+#define CONFIG_GCOV 0
+#define CONFIG_RVCT 0
+#define CONFIG_GCC 1
+#define CONFIG_MSVS 0
+#define CONFIG_PIC 1
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_CODEC_SRCS 0
+#define CONFIG_DEBUG_LIBS 0
+#define CONFIG_FAST_UNALIGNED 1
+#define CONFIG_MEM_MANAGER 0
+#define CONFIG_MEM_TRACKER 0
+#define CONFIG_MEM_CHECKS 0
+#define CONFIG_DEQUANT_TOKENS 0
+#define CONFIG_DC_RECON 0
+#define CONFIG_RUNTIME_CPU_DETECT 0
+#define CONFIG_POSTPROC 1
+#define CONFIG_VP9_POSTPROC 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_VP8_ENCODER 1
+#define CONFIG_VP8_DECODER 1
+#define CONFIG_VP9_ENCODER 0
+#define CONFIG_VP9_DECODER 1
+#define CONFIG_VP8 1
+#define CONFIG_VP9 1
+#define CONFIG_ENCODERS 1
+#define CONFIG_DECODERS 1
+#define CONFIG_STATIC_MSVCRT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_REALTIME_ONLY 0
+#define CONFIG_ONTHEFLY_BITPACKING 0
+#define CONFIG_ERROR_CONCEALMENT 0
+#define CONFIG_SHARED 0
+#define CONFIG_STATIC 1
+#define CONFIG_SMALL 0
+#define CONFIG_POSTPROC_VISUALIZER 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_UNIT_TESTS 0
+#define CONFIG_WEBM_IO 1
+#define CONFIG_DECODE_PERF_TESTS 0
+#define CONFIG_MULTI_RES_ENCODING 0
+#define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_DECRYPT 0
+#define CONFIG_MULTIPLE_ARF 0
+#define CONFIG_ALPHA 0
+#endif /* VPX_CONFIG_H */
diff --git a/x86/vpx_scale_rtcd.h b/x86/vpx_scale_rtcd.h
new file mode 100644
index 0000000..7487e5f
--- /dev/null
+++ b/x86/vpx_scale_rtcd.h
@@ -0,0 +1,69 @@
+#ifndef VPX_SCALE_RTCD_H_
+#define VPX_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct yv12_buffer_config;
+
+void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c
+
+void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c
+
+void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
+#define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c
+
+void vp8_vertical_band_2_1_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c
+
+void vp8_vertical_band_2_1_scale_i_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c
+
+void vp8_vertical_band_5_3_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c
+
+void vp8_vertical_band_5_4_scale_c(unsigned char *source, unsigned int src_pitch, unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
+#define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c
+
+void vp8_yv12_copy_frame_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vp8_yv12_copy_frame vp8_yv12_copy_frame_c
+
+void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c
+
+void vp9_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp9_extend_frame_borders vp9_extend_frame_borders_c
+
+void vp9_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf);
+#define vp9_extend_frame_inner_borders vp9_extend_frame_inner_borders_c
+
+void vpx_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
+#define vpx_yv12_copy_y vpx_yv12_copy_y_c
+
+void vpx_scale_rtcd(void);
+
+#ifdef RTCD_C
+#include "vpx_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif
diff --git a/x86/vpx_version.h b/x86/vpx_version.h
new file mode 100644
index 0000000..f60dcc7
--- /dev/null
+++ b/x86/vpx_version.h
@@ -0,0 +1,7 @@
+#define VERSION_MAJOR  1
+#define VERSION_MINOR  3
+#define VERSION_PATCH  0
+#define VERSION_EXTRA  ""
+#define VERSION_PACKED ((VERSION_MAJOR<<16)|(VERSION_MINOR<<8)|(VERSION_PATCH))
+#define VERSION_STRING_NOSP "v1.3.0"
+#define VERSION_STRING      " v1.3.0"