diff --git a/media/libvpx/update.py b/media/libvpx/update.py
index b13fd577ab2..e8112d37375 100755
--- a/media/libvpx/update.py
+++ b/media/libvpx/update.py
@@ -20,6 +20,8 @@ PLATFORMS= [
   'x86-darwin9-gcc',
   'x86_64-darwin9-gcc',
   'armv7-android-gcc',
+  'x86-win32-gcc',
+  'x86_64-win64-gcc',
 ]
 
 
diff --git a/media/libvpx/vp8_rtcd.h b/media/libvpx/vp8_rtcd.h
index 81ad9ed6dc0..e9af7505945 100644
--- a/media/libvpx/vp8_rtcd.h
+++ b/media/libvpx/vp8_rtcd.h
@@ -8,11 +8,19 @@
 
 #if defined(_WIN64)
 /* 64 bit Windows */
+#ifdef _MSC_VER
 #include "vp8_rtcd_x86_64-win64-vs8.h"
+#else
+#include "vp8_rtcd_x86_64-win64-gcc.h"
+#endif
 
 #elif defined(_WIN32)
 /* 32 bit Windows, MSVC. */
+#ifdef _MSC_VER
 #include "vp8_rtcd_x86-win32-vs8.h"
+#else
+#include "vp8_rtcd_x86-win32-gcc.h"
+#endif
 
 #elif defined(__APPLE__) && defined(__x86_64__)
 /* 64 bit MacOS. */
diff --git a/media/libvpx/vp8_rtcd_x86-win32-gcc.h b/media/libvpx/vp8_rtcd_x86-win32-gcc.h
new file mode 100644
index 00000000000..a13ca2b5a73
--- /dev/null
+++ b/media/libvpx/vp8_rtcd_x86-win32-gcc.h
@@ -0,0 +1,835 @@
+#ifndef VP8_RTCD_H_
+#define VP8_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#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_clear_system_state_c();
+void vpx_reset_mmx_state();
+RTCD_EXTERN void (*vp8_clear_system_state)();
+
+void vp8_dequantize_b_c(struct blockd*, short *dqc);
+void vp8_dequantize_b_mmx(struct blockd*, short *dqc);
+RTCD_EXTERN void (*vp8_dequantize_b)(struct blockd*, short *dqc);
+
+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);
+RTCD_EXTERN void (*vp8_dequant_idct_add)(short *input, short *dq, unsigned char *output, int stride);
+
+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);
+RTCD_EXTERN void (*vp8_dequant_idct_add_y_block)(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
+
+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);
+RTCD_EXTERN void (*vp8_dequant_idct_add_uv_block)(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
+
+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);
+RTCD_EXTERN void (*vp8_loop_filter_mbv)(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+
+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);
+RTCD_EXTERN void (*vp8_loop_filter_bv)(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+
+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);
+RTCD_EXTERN void (*vp8_loop_filter_mbh)(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+
+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);
+RTCD_EXTERN void (*vp8_loop_filter_bh)(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+
+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);
+RTCD_EXTERN void (*vp8_loop_filter_simple_mbv)(unsigned char *y, int ystride, const unsigned char *blimit);
+
+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);
+RTCD_EXTERN void (*vp8_loop_filter_simple_mbh)(unsigned char *y, int ystride, const unsigned char *blimit);
+
+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);
+RTCD_EXTERN void (*vp8_loop_filter_simple_bv)(unsigned char *y, int ystride, const unsigned char *blimit);
+
+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);
+RTCD_EXTERN void (*vp8_loop_filter_simple_bh)(unsigned char *y, int ystride, const unsigned char *blimit);
+
+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);
+RTCD_EXTERN void (*vp8_short_idct4x4llm)(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
+
+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_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);
+RTCD_EXTERN void (*vp8_short_inv_walsh4x4)(short *input, short *output);
+
+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);
+RTCD_EXTERN void (*vp8_dc_only_idct_add)(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
+
+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);
+RTCD_EXTERN void (*vp8_copy_mem16x16)(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+
+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);
+RTCD_EXTERN void (*vp8_copy_mem8x8)(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+
+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);
+RTCD_EXTERN void (*vp8_copy_mem8x4)(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+
+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);
+void vp8_build_intra_predictors_mby_s_ssse3(struct macroblockd *x, unsigned char * yabove_row, unsigned char * yleft, int left_stride, unsigned char * ypred_ptr, int y_stride);
+RTCD_EXTERN void (*vp8_build_intra_predictors_mby_s)(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_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);
+void vp8_build_intra_predictors_mbuv_s_ssse3(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);
+RTCD_EXTERN void (*vp8_build_intra_predictors_mbuv_s)(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_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_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);
+RTCD_EXTERN void (*vp8_mbpost_proc_down)(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+
+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);
+RTCD_EXTERN void (*vp8_mbpost_proc_across_ip)(unsigned char *dst, int pitch, int rows, int cols,int flimit);
+
+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);
+RTCD_EXTERN void (*vp8_post_proc_down_and_across_mb_row)(unsigned char *src, unsigned char *dst, int src_pitch, int dst_pitch, int cols, unsigned char *flimits, int size);
+
+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);
+RTCD_EXTERN void (*vp8_plane_add_noise)(unsigned char *s, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int w, unsigned int h, int pitch);
+
+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
+
+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_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);
+RTCD_EXTERN void (*vp8_filter_by_weight16x16)(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+
+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);
+RTCD_EXTERN void (*vp8_filter_by_weight8x8)(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
+
+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_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);
+void vp8_sixtap_predict16x16_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+RTCD_EXTERN void (*vp8_sixtap_predict16x16)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+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);
+void vp8_sixtap_predict8x8_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+RTCD_EXTERN void (*vp8_sixtap_predict8x8)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+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);
+void vp8_sixtap_predict8x4_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+RTCD_EXTERN void (*vp8_sixtap_predict8x4)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+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);
+void vp8_sixtap_predict4x4_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+RTCD_EXTERN void (*vp8_sixtap_predict4x4)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+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);
+void vp8_bilinear_predict16x16_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+RTCD_EXTERN void (*vp8_bilinear_predict16x16)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+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);
+void vp8_bilinear_predict8x8_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+RTCD_EXTERN void (*vp8_bilinear_predict8x8)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+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);
+RTCD_EXTERN void (*vp8_bilinear_predict8x4)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+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);
+RTCD_EXTERN void (*vp8_bilinear_predict4x4)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+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);
+RTCD_EXTERN unsigned int (*vp8_variance4x4)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+
+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);
+RTCD_EXTERN unsigned int (*vp8_variance8x8)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+
+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);
+RTCD_EXTERN unsigned int (*vp8_variance8x16)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+
+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);
+RTCD_EXTERN unsigned int (*vp8_variance16x8)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+
+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);
+RTCD_EXTERN unsigned int (*vp8_variance16x16)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+
+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);
+RTCD_EXTERN unsigned int (*vp8_sub_pixel_variance4x4)(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_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);
+RTCD_EXTERN unsigned int (*vp8_sub_pixel_variance8x8)(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_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);
+RTCD_EXTERN unsigned int (*vp8_sub_pixel_variance8x16)(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_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);
+unsigned int vp8_sub_pixel_variance16x8_ssse3(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp8_sub_pixel_variance16x8)(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_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);
+unsigned int vp8_sub_pixel_variance16x16_ssse3(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp8_sub_pixel_variance16x16)(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_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);
+RTCD_EXTERN unsigned int (*vp8_variance_halfpixvar16x16_h)(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_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);
+RTCD_EXTERN unsigned int (*vp8_variance_halfpixvar16x16_v)(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_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);
+RTCD_EXTERN unsigned int (*vp8_variance_halfpixvar16x16_hv)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+
+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);
+RTCD_EXTERN unsigned int (*vp8_sad4x4)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+
+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);
+RTCD_EXTERN unsigned int (*vp8_sad8x8)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+
+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);
+RTCD_EXTERN unsigned int (*vp8_sad8x16)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+
+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);
+RTCD_EXTERN unsigned int (*vp8_sad16x8)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+
+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);
+unsigned int vp8_sad16x16_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp8_sad16x16)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+
+void vp8_sad4x4x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp8_sad4x4x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad4x4x3)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp8_sad8x8x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp8_sad8x8x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad8x8x3)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp8_sad8x16x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp8_sad8x16x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad8x16x3)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp8_sad16x8x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp8_sad16x8x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp8_sad16x8x3_ssse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad16x8x3)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp8_sad16x16x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp8_sad16x16x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp8_sad16x16x3_ssse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad16x16x3)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp8_sad4x4x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+void vp8_sad4x4x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+RTCD_EXTERN void (*vp8_sad4x4x8)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+
+void vp8_sad8x8x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+void vp8_sad8x8x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+RTCD_EXTERN void (*vp8_sad8x8x8)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+
+void vp8_sad8x16x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+void vp8_sad8x16x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+RTCD_EXTERN void (*vp8_sad8x16x8)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+
+void vp8_sad16x8x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+void vp8_sad16x8x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+RTCD_EXTERN void (*vp8_sad16x8x8)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+
+void vp8_sad16x16x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+void vp8_sad16x16x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+RTCD_EXTERN void (*vp8_sad16x16x8)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+
+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);
+void vp8_sad4x4x4d_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad4x4x4d)(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+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);
+void vp8_sad8x8x4d_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad8x8x4d)(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+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);
+void vp8_sad8x16x4d_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad8x16x4d)(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+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);
+void vp8_sad16x8x4d_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad16x8x4d)(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+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);
+void vp8_sad16x16x4d_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad16x16x4d)(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+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 *);
+RTCD_EXTERN unsigned int (*vp8_get_mb_ss)(const short *);
+
+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);
+RTCD_EXTERN unsigned int (*vp8_sub_pixel_mse16x16)(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_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);
+RTCD_EXTERN unsigned int (*vp8_mse16x16)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
+
+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);
+RTCD_EXTERN unsigned int (*vp8_get4x4sse_cs)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
+
+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);
+void vp8_copy32xn_sse3(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, int n);
+RTCD_EXTERN void (*vp8_copy32xn)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, int n);
+
+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);
+RTCD_EXTERN void (*vp8_short_fdct4x4)(short *input, short *output, int pitch);
+
+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);
+RTCD_EXTERN void (*vp8_short_fdct8x4)(short *input, short *output, int pitch);
+
+void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
+void vp8_short_walsh4x4_sse2(short *input, short *output, int pitch);
+RTCD_EXTERN void (*vp8_short_walsh4x4)(short *input, short *output, int pitch);
+
+void vp8_regular_quantize_b_c(struct block *, struct blockd *);
+void vp8_regular_quantize_b_sse2(struct block *, struct blockd *);
+RTCD_EXTERN void (*vp8_regular_quantize_b)(struct block *, struct blockd *);
+
+void vp8_fast_quantize_b_c(struct block *, struct blockd *);
+void vp8_fast_quantize_b_sse2(struct block *, struct blockd *);
+void vp8_fast_quantize_b_ssse3(struct block *, struct blockd *);
+RTCD_EXTERN void (*vp8_fast_quantize_b)(struct block *, struct blockd *);
+
+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
+
+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_quantize_mb_c(struct macroblock *);
+#define vp8_quantize_mb vp8_quantize_mb_c
+
+void vp8_quantize_mby_c(struct macroblock *);
+#define vp8_quantize_mby vp8_quantize_mby_c
+
+void vp8_quantize_mbuv_c(struct macroblock *);
+#define vp8_quantize_mbuv vp8_quantize_mbuv_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);
+RTCD_EXTERN int (*vp8_block_error)(short *coeff, short *dqcoeff);
+
+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);
+RTCD_EXTERN int (*vp8_mbblock_error)(struct macroblock *mb, int dc);
+
+int vp8_mbuverror_c(struct macroblock *mb);
+int vp8_mbuverror_mmx(struct macroblock *mb);
+int vp8_mbuverror_xmm(struct macroblock *mb);
+RTCD_EXTERN int (*vp8_mbuverror)(struct macroblock *mb);
+
+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);
+RTCD_EXTERN void (*vp8_subtract_b)(struct block *be, struct blockd *bd, int pitch);
+
+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);
+RTCD_EXTERN void (*vp8_subtract_mby)(short *diff, unsigned char *src, int src_stride, unsigned char *pred, int pred_stride);
+
+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);
+RTCD_EXTERN void (*vp8_subtract_mbuv)(short *diff, unsigned char *usrc, unsigned char *vsrc, int src_stride, unsigned char *upred, unsigned char *vpred, int pred_stride);
+
+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);
+int vp8_full_search_sadx3(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);
+int vp8_full_search_sadx8(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);
+RTCD_EXTERN int (*vp8_full_search_sad)(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);
+
+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);
+int vp8_refining_search_sadx4(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);
+RTCD_EXTERN int (*vp8_refining_search_sad)(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);
+
+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);
+int vp8_diamond_search_sadx4(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);
+RTCD_EXTERN int (*vp8_diamond_search_sad)(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);
+
+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);
+RTCD_EXTERN void (*vp8_temporal_filter_apply)(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_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
+
+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);
+RTCD_EXTERN int (*vp8_denoiser_filter)(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);
+
+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;
+
+    vp8_clear_system_state = vp8_clear_system_state_c;
+    if (flags & HAS_MMX) vp8_clear_system_state = vpx_reset_mmx_state;
+
+    vp8_dequantize_b = vp8_dequantize_b_c;
+    if (flags & HAS_MMX) vp8_dequantize_b = vp8_dequantize_b_mmx;
+
+    vp8_dequant_idct_add = vp8_dequant_idct_add_c;
+    if (flags & HAS_MMX) vp8_dequant_idct_add = vp8_dequant_idct_add_mmx;
+
+    vp8_dequant_idct_add_y_block = vp8_dequant_idct_add_y_block_c;
+    if (flags & HAS_MMX) vp8_dequant_idct_add_y_block = vp8_dequant_idct_add_y_block_mmx;
+    if (flags & HAS_SSE2) vp8_dequant_idct_add_y_block = vp8_dequant_idct_add_y_block_sse2;
+
+    vp8_dequant_idct_add_uv_block = vp8_dequant_idct_add_uv_block_c;
+    if (flags & HAS_MMX) vp8_dequant_idct_add_uv_block = vp8_dequant_idct_add_uv_block_mmx;
+    if (flags & HAS_SSE2) vp8_dequant_idct_add_uv_block = vp8_dequant_idct_add_uv_block_sse2;
+
+    vp8_loop_filter_mbv = vp8_loop_filter_mbv_c;
+    if (flags & HAS_MMX) vp8_loop_filter_mbv = vp8_loop_filter_mbv_mmx;
+    if (flags & HAS_SSE2) vp8_loop_filter_mbv = vp8_loop_filter_mbv_sse2;
+
+    vp8_loop_filter_bv = vp8_loop_filter_bv_c;
+    if (flags & HAS_MMX) vp8_loop_filter_bv = vp8_loop_filter_bv_mmx;
+    if (flags & HAS_SSE2) vp8_loop_filter_bv = vp8_loop_filter_bv_sse2;
+
+    vp8_loop_filter_mbh = vp8_loop_filter_mbh_c;
+    if (flags & HAS_MMX) vp8_loop_filter_mbh = vp8_loop_filter_mbh_mmx;
+    if (flags & HAS_SSE2) vp8_loop_filter_mbh = vp8_loop_filter_mbh_sse2;
+
+    vp8_loop_filter_bh = vp8_loop_filter_bh_c;
+    if (flags & HAS_MMX) vp8_loop_filter_bh = vp8_loop_filter_bh_mmx;
+    if (flags & HAS_SSE2) vp8_loop_filter_bh = vp8_loop_filter_bh_sse2;
+
+    vp8_loop_filter_simple_mbv = vp8_loop_filter_simple_vertical_edge_c;
+    if (flags & HAS_MMX) vp8_loop_filter_simple_mbv = vp8_loop_filter_simple_vertical_edge_mmx;
+    if (flags & HAS_SSE2) vp8_loop_filter_simple_mbv = vp8_loop_filter_simple_vertical_edge_sse2;
+
+    vp8_loop_filter_simple_mbh = vp8_loop_filter_simple_horizontal_edge_c;
+    if (flags & HAS_MMX) vp8_loop_filter_simple_mbh = vp8_loop_filter_simple_horizontal_edge_mmx;
+    if (flags & HAS_SSE2) vp8_loop_filter_simple_mbh = vp8_loop_filter_simple_horizontal_edge_sse2;
+
+    vp8_loop_filter_simple_bv = vp8_loop_filter_bvs_c;
+    if (flags & HAS_MMX) vp8_loop_filter_simple_bv = vp8_loop_filter_bvs_mmx;
+    if (flags & HAS_SSE2) vp8_loop_filter_simple_bv = vp8_loop_filter_bvs_sse2;
+
+    vp8_loop_filter_simple_bh = vp8_loop_filter_bhs_c;
+    if (flags & HAS_MMX) vp8_loop_filter_simple_bh = vp8_loop_filter_bhs_mmx;
+    if (flags & HAS_SSE2) vp8_loop_filter_simple_bh = vp8_loop_filter_bhs_sse2;
+
+    vp8_short_idct4x4llm = vp8_short_idct4x4llm_c;
+    if (flags & HAS_MMX) vp8_short_idct4x4llm = vp8_short_idct4x4llm_mmx;
+
+
+    vp8_short_inv_walsh4x4 = vp8_short_inv_walsh4x4_c;
+    if (flags & HAS_MMX) vp8_short_inv_walsh4x4 = vp8_short_inv_walsh4x4_mmx;
+    if (flags & HAS_SSE2) vp8_short_inv_walsh4x4 = vp8_short_inv_walsh4x4_sse2;
+
+    vp8_dc_only_idct_add = vp8_dc_only_idct_add_c;
+    if (flags & HAS_MMX) vp8_dc_only_idct_add = vp8_dc_only_idct_add_mmx;
+
+    vp8_copy_mem16x16 = vp8_copy_mem16x16_c;
+    if (flags & HAS_MMX) vp8_copy_mem16x16 = vp8_copy_mem16x16_mmx;
+    if (flags & HAS_SSE2) vp8_copy_mem16x16 = vp8_copy_mem16x16_sse2;
+
+    vp8_copy_mem8x8 = vp8_copy_mem8x8_c;
+    if (flags & HAS_MMX) vp8_copy_mem8x8 = vp8_copy_mem8x8_mmx;
+
+    vp8_copy_mem8x4 = vp8_copy_mem8x4_c;
+    if (flags & HAS_MMX) vp8_copy_mem8x4 = vp8_copy_mem8x4_mmx;
+
+    vp8_build_intra_predictors_mby_s = vp8_build_intra_predictors_mby_s_c;
+    if (flags & HAS_SSE2) vp8_build_intra_predictors_mby_s = vp8_build_intra_predictors_mby_s_sse2;
+    if (flags & HAS_SSSE3) vp8_build_intra_predictors_mby_s = vp8_build_intra_predictors_mby_s_ssse3;
+
+    vp8_build_intra_predictors_mbuv_s = vp8_build_intra_predictors_mbuv_s_c;
+    if (flags & HAS_SSE2) vp8_build_intra_predictors_mbuv_s = vp8_build_intra_predictors_mbuv_s_sse2;
+    if (flags & HAS_SSSE3) vp8_build_intra_predictors_mbuv_s = vp8_build_intra_predictors_mbuv_s_ssse3;
+
+
+    vp8_mbpost_proc_down = vp8_mbpost_proc_down_c;
+    if (flags & HAS_MMX) vp8_mbpost_proc_down = vp8_mbpost_proc_down_mmx;
+    if (flags & HAS_SSE2) vp8_mbpost_proc_down = vp8_mbpost_proc_down_xmm;
+
+    vp8_mbpost_proc_across_ip = vp8_mbpost_proc_across_ip_c;
+    if (flags & HAS_SSE2) vp8_mbpost_proc_across_ip = vp8_mbpost_proc_across_ip_xmm;
+
+    vp8_post_proc_down_and_across_mb_row = vp8_post_proc_down_and_across_mb_row_c;
+    if (flags & HAS_SSE2) vp8_post_proc_down_and_across_mb_row = vp8_post_proc_down_and_across_mb_row_sse2;
+
+    vp8_plane_add_noise = vp8_plane_add_noise_c;
+    if (flags & HAS_MMX) vp8_plane_add_noise = vp8_plane_add_noise_mmx;
+    if (flags & HAS_SSE2) vp8_plane_add_noise = vp8_plane_add_noise_wmt;
+
+
+
+
+    vp8_filter_by_weight16x16 = vp8_filter_by_weight16x16_c;
+    if (flags & HAS_SSE2) vp8_filter_by_weight16x16 = vp8_filter_by_weight16x16_sse2;
+
+    vp8_filter_by_weight8x8 = vp8_filter_by_weight8x8_c;
+    if (flags & HAS_SSE2) vp8_filter_by_weight8x8 = vp8_filter_by_weight8x8_sse2;
+
+
+    vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_c;
+    if (flags & HAS_MMX) vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_mmx;
+    if (flags & HAS_SSE2) vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_sse2;
+    if (flags & HAS_SSSE3) vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_ssse3;
+
+    vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_c;
+    if (flags & HAS_MMX) vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_mmx;
+    if (flags & HAS_SSE2) vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_sse2;
+    if (flags & HAS_SSSE3) vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_ssse3;
+
+    vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_c;
+    if (flags & HAS_MMX) vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_mmx;
+    if (flags & HAS_SSE2) vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_sse2;
+    if (flags & HAS_SSSE3) vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_ssse3;
+
+    vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_c;
+    if (flags & HAS_MMX) vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_mmx;
+    if (flags & HAS_SSSE3) vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_ssse3;
+
+    vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_c;
+    if (flags & HAS_MMX) vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_mmx;
+    if (flags & HAS_SSE2) vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_sse2;
+    if (flags & HAS_SSSE3) vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_ssse3;
+
+    vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_c;
+    if (flags & HAS_MMX) vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_mmx;
+    if (flags & HAS_SSE2) vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_sse2;
+    if (flags & HAS_SSSE3) vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_ssse3;
+
+    vp8_bilinear_predict8x4 = vp8_bilinear_predict8x4_c;
+    if (flags & HAS_MMX) vp8_bilinear_predict8x4 = vp8_bilinear_predict8x4_mmx;
+
+    vp8_bilinear_predict4x4 = vp8_bilinear_predict4x4_c;
+    if (flags & HAS_MMX) vp8_bilinear_predict4x4 = vp8_bilinear_predict4x4_mmx;
+
+    vp8_variance4x4 = vp8_variance4x4_c;
+    if (flags & HAS_MMX) vp8_variance4x4 = vp8_variance4x4_mmx;
+    if (flags & HAS_SSE2) vp8_variance4x4 = vp8_variance4x4_wmt;
+
+    vp8_variance8x8 = vp8_variance8x8_c;
+    if (flags & HAS_MMX) vp8_variance8x8 = vp8_variance8x8_mmx;
+    if (flags & HAS_SSE2) vp8_variance8x8 = vp8_variance8x8_wmt;
+
+    vp8_variance8x16 = vp8_variance8x16_c;
+    if (flags & HAS_MMX) vp8_variance8x16 = vp8_variance8x16_mmx;
+    if (flags & HAS_SSE2) vp8_variance8x16 = vp8_variance8x16_wmt;
+
+    vp8_variance16x8 = vp8_variance16x8_c;
+    if (flags & HAS_MMX) vp8_variance16x8 = vp8_variance16x8_mmx;
+    if (flags & HAS_SSE2) vp8_variance16x8 = vp8_variance16x8_wmt;
+
+    vp8_variance16x16 = vp8_variance16x16_c;
+    if (flags & HAS_MMX) vp8_variance16x16 = vp8_variance16x16_mmx;
+    if (flags & HAS_SSE2) vp8_variance16x16 = vp8_variance16x16_wmt;
+
+    vp8_sub_pixel_variance4x4 = vp8_sub_pixel_variance4x4_c;
+    if (flags & HAS_MMX) vp8_sub_pixel_variance4x4 = vp8_sub_pixel_variance4x4_mmx;
+    if (flags & HAS_SSE2) vp8_sub_pixel_variance4x4 = vp8_sub_pixel_variance4x4_wmt;
+
+    vp8_sub_pixel_variance8x8 = vp8_sub_pixel_variance8x8_c;
+    if (flags & HAS_MMX) vp8_sub_pixel_variance8x8 = vp8_sub_pixel_variance8x8_mmx;
+    if (flags & HAS_SSE2) vp8_sub_pixel_variance8x8 = vp8_sub_pixel_variance8x8_wmt;
+
+    vp8_sub_pixel_variance8x16 = vp8_sub_pixel_variance8x16_c;
+    if (flags & HAS_MMX) vp8_sub_pixel_variance8x16 = vp8_sub_pixel_variance8x16_mmx;
+    if (flags & HAS_SSE2) vp8_sub_pixel_variance8x16 = vp8_sub_pixel_variance8x16_wmt;
+
+    vp8_sub_pixel_variance16x8 = vp8_sub_pixel_variance16x8_c;
+    if (flags & HAS_MMX) vp8_sub_pixel_variance16x8 = vp8_sub_pixel_variance16x8_mmx;
+    if (flags & HAS_SSE2) vp8_sub_pixel_variance16x8 = vp8_sub_pixel_variance16x8_wmt;
+    if (flags & HAS_SSSE3) vp8_sub_pixel_variance16x8 = vp8_sub_pixel_variance16x8_ssse3;
+
+    vp8_sub_pixel_variance16x16 = vp8_sub_pixel_variance16x16_c;
+    if (flags & HAS_MMX) vp8_sub_pixel_variance16x16 = vp8_sub_pixel_variance16x16_mmx;
+    if (flags & HAS_SSE2) vp8_sub_pixel_variance16x16 = vp8_sub_pixel_variance16x16_wmt;
+    if (flags & HAS_SSSE3) vp8_sub_pixel_variance16x16 = vp8_sub_pixel_variance16x16_ssse3;
+
+    vp8_variance_halfpixvar16x16_h = vp8_variance_halfpixvar16x16_h_c;
+    if (flags & HAS_MMX) vp8_variance_halfpixvar16x16_h = vp8_variance_halfpixvar16x16_h_mmx;
+    if (flags & HAS_SSE2) vp8_variance_halfpixvar16x16_h = vp8_variance_halfpixvar16x16_h_wmt;
+
+    vp8_variance_halfpixvar16x16_v = vp8_variance_halfpixvar16x16_v_c;
+    if (flags & HAS_MMX) vp8_variance_halfpixvar16x16_v = vp8_variance_halfpixvar16x16_v_mmx;
+    if (flags & HAS_SSE2) vp8_variance_halfpixvar16x16_v = vp8_variance_halfpixvar16x16_v_wmt;
+
+    vp8_variance_halfpixvar16x16_hv = vp8_variance_halfpixvar16x16_hv_c;
+    if (flags & HAS_MMX) vp8_variance_halfpixvar16x16_hv = vp8_variance_halfpixvar16x16_hv_mmx;
+    if (flags & HAS_SSE2) vp8_variance_halfpixvar16x16_hv = vp8_variance_halfpixvar16x16_hv_wmt;
+
+    vp8_sad4x4 = vp8_sad4x4_c;
+    if (flags & HAS_MMX) vp8_sad4x4 = vp8_sad4x4_mmx;
+    if (flags & HAS_SSE2) vp8_sad4x4 = vp8_sad4x4_wmt;
+
+    vp8_sad8x8 = vp8_sad8x8_c;
+    if (flags & HAS_MMX) vp8_sad8x8 = vp8_sad8x8_mmx;
+    if (flags & HAS_SSE2) vp8_sad8x8 = vp8_sad8x8_wmt;
+
+    vp8_sad8x16 = vp8_sad8x16_c;
+    if (flags & HAS_MMX) vp8_sad8x16 = vp8_sad8x16_mmx;
+    if (flags & HAS_SSE2) vp8_sad8x16 = vp8_sad8x16_wmt;
+
+    vp8_sad16x8 = vp8_sad16x8_c;
+    if (flags & HAS_MMX) vp8_sad16x8 = vp8_sad16x8_mmx;
+    if (flags & HAS_SSE2) vp8_sad16x8 = vp8_sad16x8_wmt;
+
+    vp8_sad16x16 = vp8_sad16x16_c;
+    if (flags & HAS_MMX) vp8_sad16x16 = vp8_sad16x16_mmx;
+    if (flags & HAS_SSE2) vp8_sad16x16 = vp8_sad16x16_wmt;
+    if (flags & HAS_SSE3) vp8_sad16x16 = vp8_sad16x16_sse3;
+
+    vp8_sad4x4x3 = vp8_sad4x4x3_c;
+    if (flags & HAS_SSE3) vp8_sad4x4x3 = vp8_sad4x4x3_sse3;
+
+    vp8_sad8x8x3 = vp8_sad8x8x3_c;
+    if (flags & HAS_SSE3) vp8_sad8x8x3 = vp8_sad8x8x3_sse3;
+
+    vp8_sad8x16x3 = vp8_sad8x16x3_c;
+    if (flags & HAS_SSE3) vp8_sad8x16x3 = vp8_sad8x16x3_sse3;
+
+    vp8_sad16x8x3 = vp8_sad16x8x3_c;
+    if (flags & HAS_SSE3) vp8_sad16x8x3 = vp8_sad16x8x3_sse3;
+    if (flags & HAS_SSSE3) vp8_sad16x8x3 = vp8_sad16x8x3_ssse3;
+
+    vp8_sad16x16x3 = vp8_sad16x16x3_c;
+    if (flags & HAS_SSE3) vp8_sad16x16x3 = vp8_sad16x16x3_sse3;
+    if (flags & HAS_SSSE3) vp8_sad16x16x3 = vp8_sad16x16x3_ssse3;
+
+    vp8_sad4x4x8 = vp8_sad4x4x8_c;
+    if (flags & HAS_SSE4_1) vp8_sad4x4x8 = vp8_sad4x4x8_sse4;
+
+    vp8_sad8x8x8 = vp8_sad8x8x8_c;
+    if (flags & HAS_SSE4_1) vp8_sad8x8x8 = vp8_sad8x8x8_sse4;
+
+    vp8_sad8x16x8 = vp8_sad8x16x8_c;
+    if (flags & HAS_SSE4_1) vp8_sad8x16x8 = vp8_sad8x16x8_sse4;
+
+    vp8_sad16x8x8 = vp8_sad16x8x8_c;
+    if (flags & HAS_SSE4_1) vp8_sad16x8x8 = vp8_sad16x8x8_sse4;
+
+    vp8_sad16x16x8 = vp8_sad16x16x8_c;
+    if (flags & HAS_SSE4_1) vp8_sad16x16x8 = vp8_sad16x16x8_sse4;
+
+    vp8_sad4x4x4d = vp8_sad4x4x4d_c;
+    if (flags & HAS_SSE3) vp8_sad4x4x4d = vp8_sad4x4x4d_sse3;
+
+    vp8_sad8x8x4d = vp8_sad8x8x4d_c;
+    if (flags & HAS_SSE3) vp8_sad8x8x4d = vp8_sad8x8x4d_sse3;
+
+    vp8_sad8x16x4d = vp8_sad8x16x4d_c;
+    if (flags & HAS_SSE3) vp8_sad8x16x4d = vp8_sad8x16x4d_sse3;
+
+    vp8_sad16x8x4d = vp8_sad16x8x4d_c;
+    if (flags & HAS_SSE3) vp8_sad16x8x4d = vp8_sad16x8x4d_sse3;
+
+    vp8_sad16x16x4d = vp8_sad16x16x4d_c;
+    if (flags & HAS_SSE3) vp8_sad16x16x4d = vp8_sad16x16x4d_sse3;
+
+    vp8_get_mb_ss = vp8_get_mb_ss_c;
+    if (flags & HAS_MMX) vp8_get_mb_ss = vp8_get_mb_ss_mmx;
+    if (flags & HAS_SSE2) vp8_get_mb_ss = vp8_get_mb_ss_sse2;
+
+    vp8_sub_pixel_mse16x16 = vp8_sub_pixel_mse16x16_c;
+    if (flags & HAS_MMX) vp8_sub_pixel_mse16x16 = vp8_sub_pixel_mse16x16_mmx;
+    if (flags & HAS_SSE2) vp8_sub_pixel_mse16x16 = vp8_sub_pixel_mse16x16_wmt;
+
+    vp8_mse16x16 = vp8_mse16x16_c;
+    if (flags & HAS_MMX) vp8_mse16x16 = vp8_mse16x16_mmx;
+    if (flags & HAS_SSE2) vp8_mse16x16 = vp8_mse16x16_wmt;
+
+    vp8_get4x4sse_cs = vp8_get4x4sse_cs_c;
+    if (flags & HAS_MMX) vp8_get4x4sse_cs = vp8_get4x4sse_cs_mmx;
+
+    vp8_copy32xn = vp8_copy32xn_c;
+    if (flags & HAS_SSE2) vp8_copy32xn = vp8_copy32xn_sse2;
+    if (flags & HAS_SSE3) vp8_copy32xn = vp8_copy32xn_sse3;
+
+    vp8_short_fdct4x4 = vp8_short_fdct4x4_c;
+    if (flags & HAS_MMX) vp8_short_fdct4x4 = vp8_short_fdct4x4_mmx;
+    if (flags & HAS_SSE2) vp8_short_fdct4x4 = vp8_short_fdct4x4_sse2;
+
+    vp8_short_fdct8x4 = vp8_short_fdct8x4_c;
+    if (flags & HAS_MMX) vp8_short_fdct8x4 = vp8_short_fdct8x4_mmx;
+    if (flags & HAS_SSE2) vp8_short_fdct8x4 = vp8_short_fdct8x4_sse2;
+
+    vp8_short_walsh4x4 = vp8_short_walsh4x4_c;
+    if (flags & HAS_SSE2) vp8_short_walsh4x4 = vp8_short_walsh4x4_sse2;
+
+    vp8_regular_quantize_b = vp8_regular_quantize_b_c;
+    if (flags & HAS_SSE2) vp8_regular_quantize_b = vp8_regular_quantize_b_sse2;
+
+    vp8_fast_quantize_b = vp8_fast_quantize_b_c;
+    if (flags & HAS_SSE2) vp8_fast_quantize_b = vp8_fast_quantize_b_sse2;
+    if (flags & HAS_SSSE3) vp8_fast_quantize_b = vp8_fast_quantize_b_ssse3;
+
+
+
+
+
+
+    vp8_block_error = vp8_block_error_c;
+    if (flags & HAS_MMX) vp8_block_error = vp8_block_error_mmx;
+    if (flags & HAS_SSE2) vp8_block_error = vp8_block_error_xmm;
+
+    vp8_mbblock_error = vp8_mbblock_error_c;
+    if (flags & HAS_MMX) vp8_mbblock_error = vp8_mbblock_error_mmx;
+    if (flags & HAS_SSE2) vp8_mbblock_error = vp8_mbblock_error_xmm;
+
+    vp8_mbuverror = vp8_mbuverror_c;
+    if (flags & HAS_MMX) vp8_mbuverror = vp8_mbuverror_mmx;
+    if (flags & HAS_SSE2) vp8_mbuverror = vp8_mbuverror_xmm;
+
+    vp8_subtract_b = vp8_subtract_b_c;
+    if (flags & HAS_MMX) vp8_subtract_b = vp8_subtract_b_mmx;
+    if (flags & HAS_SSE2) vp8_subtract_b = vp8_subtract_b_sse2;
+
+    vp8_subtract_mby = vp8_subtract_mby_c;
+    if (flags & HAS_MMX) vp8_subtract_mby = vp8_subtract_mby_mmx;
+    if (flags & HAS_SSE2) vp8_subtract_mby = vp8_subtract_mby_sse2;
+
+    vp8_subtract_mbuv = vp8_subtract_mbuv_c;
+    if (flags & HAS_MMX) vp8_subtract_mbuv = vp8_subtract_mbuv_mmx;
+    if (flags & HAS_SSE2) vp8_subtract_mbuv = vp8_subtract_mbuv_sse2;
+
+    vp8_full_search_sad = vp8_full_search_sad_c;
+    if (flags & HAS_SSE3) vp8_full_search_sad = vp8_full_search_sadx3;
+    if (flags & HAS_SSE4_1) vp8_full_search_sad = vp8_full_search_sadx8;
+
+    vp8_refining_search_sad = vp8_refining_search_sad_c;
+    if (flags & HAS_SSE3) vp8_refining_search_sad = vp8_refining_search_sadx4;
+
+    vp8_diamond_search_sad = vp8_diamond_search_sad_c;
+    if (flags & HAS_SSE3) vp8_diamond_search_sad = vp8_diamond_search_sadx4;
+
+    vp8_temporal_filter_apply = vp8_temporal_filter_apply_c;
+    if (flags & HAS_SSE2) vp8_temporal_filter_apply = vp8_temporal_filter_apply_sse2;
+
+
+    vp8_denoiser_filter = vp8_denoiser_filter_c;
+    if (flags & HAS_SSE2) vp8_denoiser_filter = vp8_denoiser_filter_sse2;
+}
+#endif
+#endif
diff --git a/media/libvpx/vp8_rtcd_x86_64-win64-gcc.h b/media/libvpx/vp8_rtcd_x86_64-win64-gcc.h
new file mode 100644
index 00000000000..ee5ba1c3f5c
--- /dev/null
+++ b/media/libvpx/vp8_rtcd_x86_64-win64-gcc.h
@@ -0,0 +1,655 @@
+#ifndef VP8_RTCD_H_
+#define VP8_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#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_clear_system_state_c();
+void vpx_reset_mmx_state();
+#define vp8_clear_system_state vpx_reset_mmx_state
+
+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
+
+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_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_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_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_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_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_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
+
+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_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_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_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_1_c(short *input, short *output);
+#define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
+
+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_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
+
+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_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_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_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);
+void vp8_build_intra_predictors_mby_s_ssse3(struct macroblockd *x, unsigned char * yabove_row, unsigned char * yleft, int left_stride, unsigned char * ypred_ptr, int y_stride);
+RTCD_EXTERN void (*vp8_build_intra_predictors_mby_s)(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_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);
+void vp8_build_intra_predictors_mbuv_s_ssse3(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);
+RTCD_EXTERN void (*vp8_build_intra_predictors_mbuv_s)(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_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_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
+
+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_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_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_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
+
+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_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_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
+
+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_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);
+void vp8_sixtap_predict16x16_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+RTCD_EXTERN void (*vp8_sixtap_predict16x16)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+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);
+void vp8_sixtap_predict8x8_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+RTCD_EXTERN void (*vp8_sixtap_predict8x8)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+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);
+void vp8_sixtap_predict8x4_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+RTCD_EXTERN void (*vp8_sixtap_predict8x4)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+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);
+void vp8_sixtap_predict4x4_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+RTCD_EXTERN void (*vp8_sixtap_predict4x4)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+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);
+void vp8_bilinear_predict16x16_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+RTCD_EXTERN void (*vp8_bilinear_predict16x16)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+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);
+void vp8_bilinear_predict8x8_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+RTCD_EXTERN void (*vp8_bilinear_predict8x8)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
+
+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_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
+
+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_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_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_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_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_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_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
+
+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_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);
+unsigned int vp8_sub_pixel_variance16x8_ssse3(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp8_sub_pixel_variance16x8)(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_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);
+unsigned int vp8_sub_pixel_variance16x16_ssse3(const unsigned char  *src_ptr, int  source_stride, int  xoffset, int  yoffset, const unsigned char *ref_ptr, int Refstride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp8_sub_pixel_variance16x16)(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_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_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
+
+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_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
+
+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
+
+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
+
+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
+
+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);
+unsigned int vp8_sad16x16_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp8_sad16x16)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int ref_stride, unsigned int max_sad);
+
+void vp8_sad4x4x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp8_sad4x4x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad4x4x3)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp8_sad8x8x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp8_sad8x8x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad8x8x3)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp8_sad8x16x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp8_sad8x16x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad8x16x3)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp8_sad16x8x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp8_sad16x8x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp8_sad16x8x3_ssse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad16x8x3)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp8_sad16x16x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp8_sad16x16x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp8_sad16x16x3_ssse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad16x16x3)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp8_sad4x4x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+void vp8_sad4x4x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+RTCD_EXTERN void (*vp8_sad4x4x8)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+
+void vp8_sad8x8x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+void vp8_sad8x8x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+RTCD_EXTERN void (*vp8_sad8x8x8)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+
+void vp8_sad8x16x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+void vp8_sad8x16x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+RTCD_EXTERN void (*vp8_sad8x16x8)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+
+void vp8_sad16x8x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+void vp8_sad16x8x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+RTCD_EXTERN void (*vp8_sad16x8x8)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+
+void vp8_sad16x16x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+void vp8_sad16x16x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+RTCD_EXTERN void (*vp8_sad16x16x8)(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
+
+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);
+void vp8_sad4x4x4d_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad4x4x4d)(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+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);
+void vp8_sad8x8x4d_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad8x8x4d)(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+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);
+void vp8_sad8x16x4d_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad8x16x4d)(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+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);
+void vp8_sad16x8x4d_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad16x8x4d)(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+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);
+void vp8_sad16x16x4d_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp8_sad16x16x4d)(const unsigned char *src_ptr, int src_stride, const unsigned char * const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+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
+
+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_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
+
+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
+
+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);
+void vp8_copy32xn_sse3(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, int n);
+RTCD_EXTERN void (*vp8_copy32xn)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, int n);
+
+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_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_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_fast_quantize_b_c(struct block *, struct blockd *);
+void vp8_fast_quantize_b_sse2(struct block *, struct blockd *);
+void vp8_fast_quantize_b_ssse3(struct block *, struct blockd *);
+RTCD_EXTERN void (*vp8_fast_quantize_b)(struct block *, struct blockd *);
+
+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
+
+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_quantize_mb_c(struct macroblock *);
+#define vp8_quantize_mb vp8_quantize_mb_c
+
+void vp8_quantize_mby_c(struct macroblock *);
+#define vp8_quantize_mby vp8_quantize_mby_c
+
+void vp8_quantize_mbuv_c(struct macroblock *);
+#define vp8_quantize_mbuv vp8_quantize_mbuv_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
+
+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
+
+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
+
+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_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_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
+
+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);
+int vp8_full_search_sadx3(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);
+int vp8_full_search_sadx8(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);
+RTCD_EXTERN int (*vp8_full_search_sad)(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);
+
+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);
+int vp8_refining_search_sadx4(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);
+RTCD_EXTERN int (*vp8_refining_search_sad)(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);
+
+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);
+int vp8_diamond_search_sadx4(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);
+RTCD_EXTERN int (*vp8_diamond_search_sad)(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);
+
+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
+
+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
+
+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_rtcd(void);
+
+#ifdef RTCD_C
+#include "vpx_ports/x86.h"
+static void setup_rtcd_internal(void)
+{
+    int flags = x86_simd_caps();
+
+    (void)flags;
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+    vp8_build_intra_predictors_mby_s = vp8_build_intra_predictors_mby_s_sse2;
+    if (flags & HAS_SSSE3) vp8_build_intra_predictors_mby_s = vp8_build_intra_predictors_mby_s_ssse3;
+
+    vp8_build_intra_predictors_mbuv_s = vp8_build_intra_predictors_mbuv_s_sse2;
+    if (flags & HAS_SSSE3) vp8_build_intra_predictors_mbuv_s = vp8_build_intra_predictors_mbuv_s_ssse3;
+
+
+
+
+
+
+
+
+
+
+
+
+    vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_sse2;
+    if (flags & HAS_SSSE3) vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_ssse3;
+
+    vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_sse2;
+    if (flags & HAS_SSSE3) vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_ssse3;
+
+    vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_sse2;
+    if (flags & HAS_SSSE3) vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_ssse3;
+
+    vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_mmx;
+    if (flags & HAS_SSSE3) vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_ssse3;
+
+    vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_sse2;
+    if (flags & HAS_SSSE3) vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_ssse3;
+
+    vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_sse2;
+    if (flags & HAS_SSSE3) vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_ssse3;
+
+
+
+
+
+
+
+
+
+
+
+    vp8_sub_pixel_variance16x8 = vp8_sub_pixel_variance16x8_wmt;
+    if (flags & HAS_SSSE3) vp8_sub_pixel_variance16x8 = vp8_sub_pixel_variance16x8_ssse3;
+
+    vp8_sub_pixel_variance16x16 = vp8_sub_pixel_variance16x16_wmt;
+    if (flags & HAS_SSSE3) vp8_sub_pixel_variance16x16 = vp8_sub_pixel_variance16x16_ssse3;
+
+
+
+
+
+
+
+
+    vp8_sad16x16 = vp8_sad16x16_wmt;
+    if (flags & HAS_SSE3) vp8_sad16x16 = vp8_sad16x16_sse3;
+
+    vp8_sad4x4x3 = vp8_sad4x4x3_c;
+    if (flags & HAS_SSE3) vp8_sad4x4x3 = vp8_sad4x4x3_sse3;
+
+    vp8_sad8x8x3 = vp8_sad8x8x3_c;
+    if (flags & HAS_SSE3) vp8_sad8x8x3 = vp8_sad8x8x3_sse3;
+
+    vp8_sad8x16x3 = vp8_sad8x16x3_c;
+    if (flags & HAS_SSE3) vp8_sad8x16x3 = vp8_sad8x16x3_sse3;
+
+    vp8_sad16x8x3 = vp8_sad16x8x3_c;
+    if (flags & HAS_SSE3) vp8_sad16x8x3 = vp8_sad16x8x3_sse3;
+    if (flags & HAS_SSSE3) vp8_sad16x8x3 = vp8_sad16x8x3_ssse3;
+
+    vp8_sad16x16x3 = vp8_sad16x16x3_c;
+    if (flags & HAS_SSE3) vp8_sad16x16x3 = vp8_sad16x16x3_sse3;
+    if (flags & HAS_SSSE3) vp8_sad16x16x3 = vp8_sad16x16x3_ssse3;
+
+    vp8_sad4x4x8 = vp8_sad4x4x8_c;
+    if (flags & HAS_SSE4_1) vp8_sad4x4x8 = vp8_sad4x4x8_sse4;
+
+    vp8_sad8x8x8 = vp8_sad8x8x8_c;
+    if (flags & HAS_SSE4_1) vp8_sad8x8x8 = vp8_sad8x8x8_sse4;
+
+    vp8_sad8x16x8 = vp8_sad8x16x8_c;
+    if (flags & HAS_SSE4_1) vp8_sad8x16x8 = vp8_sad8x16x8_sse4;
+
+    vp8_sad16x8x8 = vp8_sad16x8x8_c;
+    if (flags & HAS_SSE4_1) vp8_sad16x8x8 = vp8_sad16x8x8_sse4;
+
+    vp8_sad16x16x8 = vp8_sad16x16x8_c;
+    if (flags & HAS_SSE4_1) vp8_sad16x16x8 = vp8_sad16x16x8_sse4;
+
+    vp8_sad4x4x4d = vp8_sad4x4x4d_c;
+    if (flags & HAS_SSE3) vp8_sad4x4x4d = vp8_sad4x4x4d_sse3;
+
+    vp8_sad8x8x4d = vp8_sad8x8x4d_c;
+    if (flags & HAS_SSE3) vp8_sad8x8x4d = vp8_sad8x8x4d_sse3;
+
+    vp8_sad8x16x4d = vp8_sad8x16x4d_c;
+    if (flags & HAS_SSE3) vp8_sad8x16x4d = vp8_sad8x16x4d_sse3;
+
+    vp8_sad16x8x4d = vp8_sad16x8x4d_c;
+    if (flags & HAS_SSE3) vp8_sad16x8x4d = vp8_sad16x8x4d_sse3;
+
+    vp8_sad16x16x4d = vp8_sad16x16x4d_c;
+    if (flags & HAS_SSE3) vp8_sad16x16x4d = vp8_sad16x16x4d_sse3;
+
+
+
+
+
+    vp8_copy32xn = vp8_copy32xn_sse2;
+    if (flags & HAS_SSE3) vp8_copy32xn = vp8_copy32xn_sse3;
+
+
+
+
+
+    vp8_fast_quantize_b = vp8_fast_quantize_b_sse2;
+    if (flags & HAS_SSSE3) vp8_fast_quantize_b = vp8_fast_quantize_b_ssse3;
+
+
+
+
+
+
+
+
+
+
+
+
+    vp8_full_search_sad = vp8_full_search_sad_c;
+    if (flags & HAS_SSE3) vp8_full_search_sad = vp8_full_search_sadx3;
+    if (flags & HAS_SSE4_1) vp8_full_search_sad = vp8_full_search_sadx8;
+
+    vp8_refining_search_sad = vp8_refining_search_sad_c;
+    if (flags & HAS_SSE3) vp8_refining_search_sad = vp8_refining_search_sadx4;
+
+    vp8_diamond_search_sad = vp8_diamond_search_sad_c;
+    if (flags & HAS_SSE3) vp8_diamond_search_sad = vp8_diamond_search_sadx4;
+}
+#endif
+#endif
diff --git a/media/libvpx/vp9_rtcd.h b/media/libvpx/vp9_rtcd.h
index 49f1ce3a8a7..21ad536010d 100644
--- a/media/libvpx/vp9_rtcd.h
+++ b/media/libvpx/vp9_rtcd.h
@@ -8,11 +8,19 @@
 
 #if defined(_WIN64)
 /* 64 bit Windows */
+#ifdef _MSC_VER
 #include "vp9_rtcd_x86_64-win64-vs8.h"
+#else
+#include "vp9_rtcd_x86_64-win64-gcc.h"
+#endif
 
 #elif defined(_WIN32)
 /* 32 bit Windows, MSVC. */
+#ifdef _MSC_VER
 #include "vp9_rtcd_x86-win32-vs8.h"
+#else
+#include "vp9_rtcd_x86-win32-gcc.h"
+#endif
 
 #elif defined(__APPLE__) && defined(__x86_64__)
 /* 64 bit MacOS. */
diff --git a/media/libvpx/vp9_rtcd_x86-win32-gcc.h b/media/libvpx/vp9_rtcd_x86-win32-gcc.h
new file mode 100644
index 00000000000..ba86228db6a
--- /dev/null
+++ b/media/libvpx/vp9_rtcd_x86-win32-gcc.h
@@ -0,0 +1,1470 @@
+#ifndef VP9_RTCD_H_
+#define VP9_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#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]
+union int_mv;
+struct yv12_buffer_config;
+
+void vp9_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d207_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d207_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d45_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d45_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d63_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d63_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_h_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d153_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d153_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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);
+RTCD_EXTERN void (*vp9_v_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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);
+RTCD_EXTERN void (*vp9_tm_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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);
+RTCD_EXTERN void (*vp9_dc_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_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_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d207_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d207_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d45_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d45_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d63_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d63_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_h_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d153_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d153_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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);
+RTCD_EXTERN void (*vp9_v_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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);
+RTCD_EXTERN void (*vp9_tm_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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);
+RTCD_EXTERN void (*vp9_dc_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_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_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d207_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d207_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d45_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d45_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d63_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d63_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d153_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d153_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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);
+RTCD_EXTERN void (*vp9_v_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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);
+RTCD_EXTERN void (*vp9_tm_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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);
+RTCD_EXTERN void (*vp9_dc_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_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_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d207_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d207_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d45_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d45_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d63_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d63_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_h_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_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_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);
+RTCD_EXTERN void (*vp9_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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);
+RTCD_EXTERN void (*vp9_dc_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_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_mb_lpf_vertical_edge_w_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vp9_mb_lpf_vertical_edge_w_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+RTCD_EXTERN void (*vp9_mb_lpf_vertical_edge_w)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+
+void vp9_mbloop_filter_vertical_edge_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_mbloop_filter_vertical_edge_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+RTCD_EXTERN void (*vp9_mbloop_filter_vertical_edge)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+
+void vp9_loop_filter_vertical_edge_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_loop_filter_vertical_edge_mmx(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+RTCD_EXTERN void (*vp9_loop_filter_vertical_edge)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+
+void vp9_mb_lpf_horizontal_edge_w_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_mb_lpf_horizontal_edge_w_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+RTCD_EXTERN void (*vp9_mb_lpf_horizontal_edge_w)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+
+void vp9_mbloop_filter_horizontal_edge_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_mbloop_filter_horizontal_edge_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+RTCD_EXTERN void (*vp9_mbloop_filter_horizontal_edge)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+
+void vp9_loop_filter_horizontal_edge_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_loop_filter_horizontal_edge_mmx(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+RTCD_EXTERN void (*vp9_loop_filter_horizontal_edge)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+
+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_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_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);
+RTCD_EXTERN void (*vp9_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 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);
+RTCD_EXTERN void (*vp9_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 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);
+void vp9_convolve8_ssse3(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 (*vp9_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 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);
+void vp9_convolve8_horiz_ssse3(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 (*vp9_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 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);
+void vp9_convolve8_vert_ssse3(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 (*vp9_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 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);
+void vp9_convolve8_avg_ssse3(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 (*vp9_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 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);
+void vp9_convolve8_avg_horiz_ssse3(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 (*vp9_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 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);
+void vp9_convolve8_avg_vert_ssse3(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 (*vp9_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 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);
+RTCD_EXTERN void (*vp9_idct4x4_1_add)(const int16_t *input, uint8_t *dest, int dest_stride);
+
+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);
+RTCD_EXTERN void (*vp9_idct4x4_16_add)(const int16_t *input, uint8_t *dest, int dest_stride);
+
+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);
+RTCD_EXTERN void (*vp9_idct8x8_1_add)(const int16_t *input, uint8_t *dest, int dest_stride);
+
+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);
+RTCD_EXTERN void (*vp9_idct8x8_64_add)(const int16_t *input, uint8_t *dest, int dest_stride);
+
+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);
+RTCD_EXTERN void (*vp9_idct8x8_10_add)(const int16_t *input, uint8_t *dest, int dest_stride);
+
+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);
+RTCD_EXTERN void (*vp9_idct16x16_1_add)(const int16_t *input, uint8_t *dest, int dest_stride);
+
+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);
+RTCD_EXTERN void (*vp9_idct16x16_256_add)(const int16_t *input, uint8_t *dest, int dest_stride);
+
+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);
+RTCD_EXTERN void (*vp9_idct16x16_10_add)(const int16_t *input, uint8_t *dest, int dest_stride);
+
+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);
+RTCD_EXTERN void (*vp9_idct32x32_1024_add)(const int16_t *input, uint8_t *dest, int dest_stride);
+
+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);
+RTCD_EXTERN void (*vp9_idct32x32_34_add)(const int16_t *input, uint8_t *dest, int dest_stride);
+
+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);
+RTCD_EXTERN void (*vp9_idct32x32_1_add)(const int16_t *input, uint8_t *dest, int dest_stride);
+
+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);
+RTCD_EXTERN void (*vp9_iht4x4_16_add)(const int16_t *input, uint8_t *dest, int dest_stride, int tx_type);
+
+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);
+RTCD_EXTERN void (*vp9_iht8x8_64_add)(const int16_t *input, uint8_t *dest, int dest_stride, int tx_type);
+
+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);
+RTCD_EXTERN void (*vp9_iht16x16_256_add)(const int16_t *input, uint8_t *output, int pitch, int tx_type);
+
+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_iwht4x4_16_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_iwht4x4_16_add vp9_iwht4x4_16_add_c
+
+unsigned int vp9_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x8_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x8_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+void vp9_get_sse_sum_8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vp9_get8x8var_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+RTCD_EXTERN void (*vp9_get_sse_sum_8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+
+unsigned int vp9_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance8x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance4x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance4x4_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance4x8_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance4x8_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance4x4_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance4x4_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sad64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad64x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+
+unsigned int vp9_sad32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad32x64)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+
+unsigned int vp9_sad64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad64x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+
+unsigned int vp9_sad32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad32x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+
+unsigned int vp9_sad16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad16x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+
+unsigned int vp9_sad32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad32x32)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+
+unsigned int vp9_sad16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad16x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+
+unsigned int vp9_sad16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x8_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad16x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+
+unsigned int vp9_sad8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad8x16)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+
+unsigned int vp9_sad8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x8_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad8x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+
+unsigned int vp9_sad8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad8x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+
+unsigned int vp9_sad4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x8_sse(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad4x8)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+
+unsigned int vp9_sad4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x4_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x4_sse(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad4x4)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+
+unsigned int vp9_sad64x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad64x64_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad64x64_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+
+unsigned int vp9_sad32x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x64_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad32x64_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+
+unsigned int vp9_sad64x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad64x32_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad64x32_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+
+unsigned int vp9_sad32x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x16_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad32x16_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+
+unsigned int vp9_sad16x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x32_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad16x32_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+
+unsigned int vp9_sad32x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x32_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad32x32_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+
+unsigned int vp9_sad16x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x16_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad16x16_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+
+unsigned int vp9_sad16x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x8_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad16x8_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+
+unsigned int vp9_sad8x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x16_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad8x16_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+
+unsigned int vp9_sad8x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x8_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad8x8_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+
+unsigned int vp9_sad8x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x4_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad8x4_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+
+unsigned int vp9_sad4x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad4x8_avg_sse(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad4x8_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+
+unsigned int vp9_sad4x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad4x4_avg_sse(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+RTCD_EXTERN unsigned int (*vp9_sad4x4_avg)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+
+unsigned int vp9_variance_halfpixvar16x16_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance_halfpixvar16x16_h_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance_halfpixvar16x16_h)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_variance_halfpixvar16x16_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance_halfpixvar16x16_v_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance_halfpixvar16x16_v)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_variance_halfpixvar16x16_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance_halfpixvar16x16_hv_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_variance_halfpixvar16x16_hv)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_variance_halfpixvar64x64_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar64x64_h vp9_variance_halfpixvar64x64_h_c
+
+unsigned int vp9_variance_halfpixvar64x64_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar64x64_v vp9_variance_halfpixvar64x64_v_c
+
+unsigned int vp9_variance_halfpixvar64x64_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar64x64_hv vp9_variance_halfpixvar64x64_hv_c
+
+unsigned int vp9_variance_halfpixvar32x32_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar32x32_h vp9_variance_halfpixvar32x32_h_c
+
+unsigned int vp9_variance_halfpixvar32x32_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar32x32_v vp9_variance_halfpixvar32x32_v_c
+
+unsigned int vp9_variance_halfpixvar32x32_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar32x32_hv vp9_variance_halfpixvar32x32_hv_c
+
+void vp9_sad64x64x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp9_sad64x64x3 vp9_sad64x64x3_c
+
+void vp9_sad32x32x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp9_sad32x32x3 vp9_sad32x32x3_c
+
+void vp9_sad16x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x16x3_sse3(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x16x3_ssse3(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad16x16x3)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad16x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x8x3_sse3(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x8x3_ssse3(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad16x8x3)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad8x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp9_sad8x16x3_sse3(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad8x16x3)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad8x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp9_sad8x8x3_sse3(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad8x8x3)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad4x4x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp9_sad4x4x3_sse3(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad4x4x3)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad64x64x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad64x64x8 vp9_sad64x64x8_c
+
+void vp9_sad32x32x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad32x32x8 vp9_sad32x32x8_c
+
+void vp9_sad16x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad16x16x8 vp9_sad16x16x8_c
+
+void vp9_sad16x8x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad16x8x8 vp9_sad16x8x8_c
+
+void vp9_sad8x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad8x16x8 vp9_sad8x16x8_c
+
+void vp9_sad8x8x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad8x8x8 vp9_sad8x8x8_c
+
+void vp9_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vp9_sad8x4x8 vp9_sad8x4x8_c
+
+void vp9_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vp9_sad4x8x8 vp9_sad4x8x8_c
+
+void vp9_sad4x4x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad4x4x8 vp9_sad4x4x8_c
+
+void vp9_sad64x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad64x64x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad64x64x4d)(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad32x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad32x64x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad32x64x4d)(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad64x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad64x32x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad64x32x4d)(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad32x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad32x16x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad32x16x4d)(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad16x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x32x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad16x32x4d)(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad32x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad32x32x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad32x32x4d)(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad16x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x16x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad16x16x4d)(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad16x8x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x8x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad16x8x4d)(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad8x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad8x16x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad8x16x4d)(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad8x8x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad8x8x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad8x8x4d)(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
+void vp9_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad8x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
+
+void vp9_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
+void vp9_sad4x8x4d_sse(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad4x8x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
+
+void vp9_sad4x4x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad4x4x4d_sse(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad4x4x4d)(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+
+unsigned int vp9_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vp9_mse16x16_mmx(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vp9_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+
+unsigned int vp9_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vp9_mse8x16 vp9_mse8x16_c
+
+unsigned int vp9_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vp9_mse16x8 vp9_mse16x8_c
+
+unsigned int vp9_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vp9_mse8x8 vp9_mse8x8_c
+
+unsigned int vp9_sub_pixel_mse64x64_c(const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_mse64x64 vp9_sub_pixel_mse64x64_c
+
+unsigned int vp9_sub_pixel_mse32x32_c(const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_mse32x32 vp9_sub_pixel_mse32x32_c
+
+unsigned int vp9_get_mb_ss_c(const int16_t *);
+unsigned int vp9_get_mb_ss_mmx(const int16_t *);
+unsigned int vp9_get_mb_ss_sse2(const int16_t *);
+RTCD_EXTERN unsigned int (*vp9_get_mb_ss)(const int16_t *);
+
+int64_t vp9_block_error_c(int16_t *coeff, int16_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t vp9_block_error_sse2(int16_t *coeff, int16_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+RTCD_EXTERN int64_t (*vp9_block_error)(int16_t *coeff, int16_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+
+void vp9_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);
+void vp9_subtract_block_sse2(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);
+RTCD_EXTERN void (*vp9_subtract_block)(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);
+
+void vp9_quantize_b_c(const int16_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, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_b vp9_quantize_b_c
+
+void vp9_quantize_b_32x32_c(const int16_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, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+#define vp9_quantize_b_32x32 vp9_quantize_b_32x32_c
+
+void vp9_short_fht4x4_c(const int16_t *input, int16_t *output, int stride, int tx_type);
+void vp9_short_fht4x4_sse2(const int16_t *input, int16_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*vp9_short_fht4x4)(const int16_t *input, int16_t *output, int stride, int tx_type);
+
+void vp9_short_fht8x8_c(const int16_t *input, int16_t *output, int stride, int tx_type);
+void vp9_short_fht8x8_sse2(const int16_t *input, int16_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*vp9_short_fht8x8)(const int16_t *input, int16_t *output, int stride, int tx_type);
+
+void vp9_short_fht16x16_c(const int16_t *input, int16_t *output, int stride, int tx_type);
+void vp9_short_fht16x16_sse2(const int16_t *input, int16_t *output, int stride, int tx_type);
+RTCD_EXTERN void (*vp9_short_fht16x16)(const int16_t *input, int16_t *output, int stride, int tx_type);
+
+void vp9_fwht4x4_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fwht4x4 vp9_fwht4x4_c
+
+void vp9_fdct4x4_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct4x4_sse2(const int16_t *input, int16_t *output, int stride);
+RTCD_EXTERN void (*vp9_fdct4x4)(const int16_t *input, int16_t *output, int stride);
+
+void vp9_fdct8x8_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct8x8_sse2(const int16_t *input, int16_t *output, int stride);
+RTCD_EXTERN void (*vp9_fdct8x8)(const int16_t *input, int16_t *output, int stride);
+
+void vp9_fdct16x16_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride);
+RTCD_EXTERN void (*vp9_fdct16x16)(const int16_t *input, int16_t *output, int stride);
+
+void vp9_fdct32x32_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct32x32_sse2(const int16_t *input, int16_t *output, int stride);
+RTCD_EXTERN void (*vp9_fdct32x32)(const int16_t *input, int16_t *output, int stride);
+
+void vp9_fdct32x32_rd_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct32x32_rd_sse2(const int16_t *input, int16_t *output, int stride);
+RTCD_EXTERN void (*vp9_fdct32x32_rd)(const int16_t *input, int16_t *output, int stride);
+
+int vp9_full_search_sad_c(struct macroblock *x, union int_mv *ref_mv, int sad_per_bit, int distance, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv, int n);
+int vp9_full_search_sadx3(struct macroblock *x, union int_mv *ref_mv, int sad_per_bit, int distance, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv, int n);
+int vp9_full_search_sadx8(struct macroblock *x, union int_mv *ref_mv, int sad_per_bit, int distance, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv, int n);
+RTCD_EXTERN int (*vp9_full_search_sad)(struct macroblock *x, union int_mv *ref_mv, int sad_per_bit, int distance, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv, int n);
+
+int vp9_refining_search_sad_c(struct macroblock *x, union int_mv *ref_mv, int sad_per_bit, int distance, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv);
+int vp9_refining_search_sadx4(struct macroblock *x, union int_mv *ref_mv, int sad_per_bit, int distance, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv);
+RTCD_EXTERN int (*vp9_refining_search_sad)(struct macroblock *x, union int_mv *ref_mv, int sad_per_bit, int distance, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv);
+
+int vp9_diamond_search_sad_c(struct macroblock *x, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv);
+int vp9_diamond_search_sadx4(struct macroblock *x, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv);
+RTCD_EXTERN int (*vp9_diamond_search_sad)(struct macroblock *x, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv);
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+void vp9_temporal_filter_apply_sse2(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+RTCD_EXTERN void (*vp9_temporal_filter_apply)(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+
+void vp9_yv12_copy_partial_frame_c(struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int fraction);
+#define vp9_yv12_copy_partial_frame vp9_yv12_copy_partial_frame_c
+
+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;
+
+    vp9_d207_predictor_4x4 = vp9_d207_predictor_4x4_c;
+    if (flags & HAS_SSSE3) vp9_d207_predictor_4x4 = vp9_d207_predictor_4x4_ssse3;
+
+    vp9_d45_predictor_4x4 = vp9_d45_predictor_4x4_c;
+    if (flags & HAS_SSSE3) vp9_d45_predictor_4x4 = vp9_d45_predictor_4x4_ssse3;
+
+    vp9_d63_predictor_4x4 = vp9_d63_predictor_4x4_c;
+    if (flags & HAS_SSSE3) vp9_d63_predictor_4x4 = vp9_d63_predictor_4x4_ssse3;
+
+    vp9_h_predictor_4x4 = vp9_h_predictor_4x4_c;
+    if (flags & HAS_SSSE3) vp9_h_predictor_4x4 = vp9_h_predictor_4x4_ssse3;
+
+
+
+    vp9_d153_predictor_4x4 = vp9_d153_predictor_4x4_c;
+    if (flags & HAS_SSSE3) vp9_d153_predictor_4x4 = vp9_d153_predictor_4x4_ssse3;
+
+    vp9_v_predictor_4x4 = vp9_v_predictor_4x4_c;
+    if (flags & HAS_SSE) vp9_v_predictor_4x4 = vp9_v_predictor_4x4_sse;
+
+    vp9_tm_predictor_4x4 = vp9_tm_predictor_4x4_c;
+    if (flags & HAS_SSE) vp9_tm_predictor_4x4 = vp9_tm_predictor_4x4_sse;
+
+    vp9_dc_predictor_4x4 = vp9_dc_predictor_4x4_c;
+    if (flags & HAS_SSE) vp9_dc_predictor_4x4 = vp9_dc_predictor_4x4_sse;
+
+
+
+
+    vp9_d207_predictor_8x8 = vp9_d207_predictor_8x8_c;
+    if (flags & HAS_SSSE3) vp9_d207_predictor_8x8 = vp9_d207_predictor_8x8_ssse3;
+
+    vp9_d45_predictor_8x8 = vp9_d45_predictor_8x8_c;
+    if (flags & HAS_SSSE3) vp9_d45_predictor_8x8 = vp9_d45_predictor_8x8_ssse3;
+
+    vp9_d63_predictor_8x8 = vp9_d63_predictor_8x8_c;
+    if (flags & HAS_SSSE3) vp9_d63_predictor_8x8 = vp9_d63_predictor_8x8_ssse3;
+
+    vp9_h_predictor_8x8 = vp9_h_predictor_8x8_c;
+    if (flags & HAS_SSSE3) vp9_h_predictor_8x8 = vp9_h_predictor_8x8_ssse3;
+
+
+
+    vp9_d153_predictor_8x8 = vp9_d153_predictor_8x8_c;
+    if (flags & HAS_SSSE3) vp9_d153_predictor_8x8 = vp9_d153_predictor_8x8_ssse3;
+
+    vp9_v_predictor_8x8 = vp9_v_predictor_8x8_c;
+    if (flags & HAS_SSE) vp9_v_predictor_8x8 = vp9_v_predictor_8x8_sse;
+
+    vp9_tm_predictor_8x8 = vp9_tm_predictor_8x8_c;
+    if (flags & HAS_SSE2) vp9_tm_predictor_8x8 = vp9_tm_predictor_8x8_sse2;
+
+    vp9_dc_predictor_8x8 = vp9_dc_predictor_8x8_c;
+    if (flags & HAS_SSE) vp9_dc_predictor_8x8 = vp9_dc_predictor_8x8_sse;
+
+
+
+
+    vp9_d207_predictor_16x16 = vp9_d207_predictor_16x16_c;
+    if (flags & HAS_SSSE3) vp9_d207_predictor_16x16 = vp9_d207_predictor_16x16_ssse3;
+
+    vp9_d45_predictor_16x16 = vp9_d45_predictor_16x16_c;
+    if (flags & HAS_SSSE3) vp9_d45_predictor_16x16 = vp9_d45_predictor_16x16_ssse3;
+
+    vp9_d63_predictor_16x16 = vp9_d63_predictor_16x16_c;
+    if (flags & HAS_SSSE3) vp9_d63_predictor_16x16 = vp9_d63_predictor_16x16_ssse3;
+
+    vp9_h_predictor_16x16 = vp9_h_predictor_16x16_c;
+    if (flags & HAS_SSSE3) vp9_h_predictor_16x16 = vp9_h_predictor_16x16_ssse3;
+
+
+
+    vp9_d153_predictor_16x16 = vp9_d153_predictor_16x16_c;
+    if (flags & HAS_SSSE3) vp9_d153_predictor_16x16 = vp9_d153_predictor_16x16_ssse3;
+
+    vp9_v_predictor_16x16 = vp9_v_predictor_16x16_c;
+    if (flags & HAS_SSE2) vp9_v_predictor_16x16 = vp9_v_predictor_16x16_sse2;
+
+    vp9_tm_predictor_16x16 = vp9_tm_predictor_16x16_c;
+    if (flags & HAS_SSE2) vp9_tm_predictor_16x16 = vp9_tm_predictor_16x16_sse2;
+
+    vp9_dc_predictor_16x16 = vp9_dc_predictor_16x16_c;
+    if (flags & HAS_SSE2) vp9_dc_predictor_16x16 = vp9_dc_predictor_16x16_sse2;
+
+
+
+
+    vp9_d207_predictor_32x32 = vp9_d207_predictor_32x32_c;
+    if (flags & HAS_SSSE3) vp9_d207_predictor_32x32 = vp9_d207_predictor_32x32_ssse3;
+
+    vp9_d45_predictor_32x32 = vp9_d45_predictor_32x32_c;
+    if (flags & HAS_SSSE3) vp9_d45_predictor_32x32 = vp9_d45_predictor_32x32_ssse3;
+
+    vp9_d63_predictor_32x32 = vp9_d63_predictor_32x32_c;
+    if (flags & HAS_SSSE3) vp9_d63_predictor_32x32 = vp9_d63_predictor_32x32_ssse3;
+
+    vp9_h_predictor_32x32 = vp9_h_predictor_32x32_c;
+    if (flags & HAS_SSSE3) vp9_h_predictor_32x32 = vp9_h_predictor_32x32_ssse3;
+
+
+
+
+    vp9_v_predictor_32x32 = vp9_v_predictor_32x32_c;
+    if (flags & HAS_SSE2) vp9_v_predictor_32x32 = vp9_v_predictor_32x32_sse2;
+
+
+    vp9_dc_predictor_32x32 = vp9_dc_predictor_32x32_c;
+    if (flags & HAS_SSE2) vp9_dc_predictor_32x32 = vp9_dc_predictor_32x32_sse2;
+
+
+
+
+    vp9_mb_lpf_vertical_edge_w = vp9_mb_lpf_vertical_edge_w_c;
+    if (flags & HAS_SSE2) vp9_mb_lpf_vertical_edge_w = vp9_mb_lpf_vertical_edge_w_sse2;
+
+    vp9_mbloop_filter_vertical_edge = vp9_mbloop_filter_vertical_edge_c;
+    if (flags & HAS_SSE2) vp9_mbloop_filter_vertical_edge = vp9_mbloop_filter_vertical_edge_sse2;
+
+    vp9_loop_filter_vertical_edge = vp9_loop_filter_vertical_edge_c;
+    if (flags & HAS_MMX) vp9_loop_filter_vertical_edge = vp9_loop_filter_vertical_edge_mmx;
+
+    vp9_mb_lpf_horizontal_edge_w = vp9_mb_lpf_horizontal_edge_w_c;
+    if (flags & HAS_SSE2) vp9_mb_lpf_horizontal_edge_w = vp9_mb_lpf_horizontal_edge_w_sse2;
+
+    vp9_mbloop_filter_horizontal_edge = vp9_mbloop_filter_horizontal_edge_c;
+    if (flags & HAS_SSE2) vp9_mbloop_filter_horizontal_edge = vp9_mbloop_filter_horizontal_edge_sse2;
+
+    vp9_loop_filter_horizontal_edge = vp9_loop_filter_horizontal_edge_c;
+    if (flags & HAS_MMX) vp9_loop_filter_horizontal_edge = vp9_loop_filter_horizontal_edge_mmx;
+
+
+
+
+    vp9_convolve_copy = vp9_convolve_copy_c;
+    if (flags & HAS_SSE2) vp9_convolve_copy = vp9_convolve_copy_sse2;
+
+    vp9_convolve_avg = vp9_convolve_avg_c;
+    if (flags & HAS_SSE2) vp9_convolve_avg = vp9_convolve_avg_sse2;
+
+    vp9_convolve8 = vp9_convolve8_c;
+    if (flags & HAS_SSE2) vp9_convolve8 = vp9_convolve8_sse2;
+    if (flags & HAS_SSSE3) vp9_convolve8 = vp9_convolve8_ssse3;
+
+    vp9_convolve8_horiz = vp9_convolve8_horiz_c;
+    if (flags & HAS_SSE2) vp9_convolve8_horiz = vp9_convolve8_horiz_sse2;
+    if (flags & HAS_SSSE3) vp9_convolve8_horiz = vp9_convolve8_horiz_ssse3;
+
+    vp9_convolve8_vert = vp9_convolve8_vert_c;
+    if (flags & HAS_SSE2) vp9_convolve8_vert = vp9_convolve8_vert_sse2;
+    if (flags & HAS_SSSE3) vp9_convolve8_vert = vp9_convolve8_vert_ssse3;
+
+    vp9_convolve8_avg = vp9_convolve8_avg_c;
+    if (flags & HAS_SSE2) vp9_convolve8_avg = vp9_convolve8_avg_sse2;
+    if (flags & HAS_SSSE3) vp9_convolve8_avg = vp9_convolve8_avg_ssse3;
+
+    vp9_convolve8_avg_horiz = vp9_convolve8_avg_horiz_c;
+    if (flags & HAS_SSE2) vp9_convolve8_avg_horiz = vp9_convolve8_avg_horiz_sse2;
+    if (flags & HAS_SSSE3) vp9_convolve8_avg_horiz = vp9_convolve8_avg_horiz_ssse3;
+
+    vp9_convolve8_avg_vert = vp9_convolve8_avg_vert_c;
+    if (flags & HAS_SSE2) vp9_convolve8_avg_vert = vp9_convolve8_avg_vert_sse2;
+    if (flags & HAS_SSSE3) vp9_convolve8_avg_vert = vp9_convolve8_avg_vert_ssse3;
+
+    vp9_idct4x4_1_add = vp9_idct4x4_1_add_c;
+    if (flags & HAS_SSE2) vp9_idct4x4_1_add = vp9_idct4x4_1_add_sse2;
+
+    vp9_idct4x4_16_add = vp9_idct4x4_16_add_c;
+    if (flags & HAS_SSE2) vp9_idct4x4_16_add = vp9_idct4x4_16_add_sse2;
+
+    vp9_idct8x8_1_add = vp9_idct8x8_1_add_c;
+    if (flags & HAS_SSE2) vp9_idct8x8_1_add = vp9_idct8x8_1_add_sse2;
+
+    vp9_idct8x8_64_add = vp9_idct8x8_64_add_c;
+    if (flags & HAS_SSE2) vp9_idct8x8_64_add = vp9_idct8x8_64_add_sse2;
+
+    vp9_idct8x8_10_add = vp9_idct8x8_10_add_c;
+    if (flags & HAS_SSE2) vp9_idct8x8_10_add = vp9_idct8x8_10_add_sse2;
+
+    vp9_idct16x16_1_add = vp9_idct16x16_1_add_c;
+    if (flags & HAS_SSE2) vp9_idct16x16_1_add = vp9_idct16x16_1_add_sse2;
+
+    vp9_idct16x16_256_add = vp9_idct16x16_256_add_c;
+    if (flags & HAS_SSE2) vp9_idct16x16_256_add = vp9_idct16x16_256_add_sse2;
+
+    vp9_idct16x16_10_add = vp9_idct16x16_10_add_c;
+    if (flags & HAS_SSE2) vp9_idct16x16_10_add = vp9_idct16x16_10_add_sse2;
+
+    vp9_idct32x32_1024_add = vp9_idct32x32_1024_add_c;
+    if (flags & HAS_SSE2) vp9_idct32x32_1024_add = vp9_idct32x32_1024_add_sse2;
+
+    vp9_idct32x32_34_add = vp9_idct32x32_34_add_c;
+    if (flags & HAS_SSE2) vp9_idct32x32_34_add = vp9_idct32x32_34_add_sse2;
+
+    vp9_idct32x32_1_add = vp9_idct32x32_1_add_c;
+    if (flags & HAS_SSE2) vp9_idct32x32_1_add = vp9_idct32x32_1_add_sse2;
+
+    vp9_iht4x4_16_add = vp9_iht4x4_16_add_c;
+    if (flags & HAS_SSE2) vp9_iht4x4_16_add = vp9_iht4x4_16_add_sse2;
+
+    vp9_iht8x8_64_add = vp9_iht8x8_64_add_c;
+    if (flags & HAS_SSE2) vp9_iht8x8_64_add = vp9_iht8x8_64_add_sse2;
+
+    vp9_iht16x16_256_add = vp9_iht16x16_256_add_c;
+    if (flags & HAS_SSE2) vp9_iht16x16_256_add = vp9_iht16x16_256_add_sse2;
+
+
+
+    vp9_variance32x16 = vp9_variance32x16_c;
+    if (flags & HAS_SSE2) vp9_variance32x16 = vp9_variance32x16_sse2;
+
+    vp9_variance16x32 = vp9_variance16x32_c;
+    if (flags & HAS_SSE2) vp9_variance16x32 = vp9_variance16x32_sse2;
+
+    vp9_variance64x32 = vp9_variance64x32_c;
+    if (flags & HAS_SSE2) vp9_variance64x32 = vp9_variance64x32_sse2;
+
+    vp9_variance32x64 = vp9_variance32x64_c;
+    if (flags & HAS_SSE2) vp9_variance32x64 = vp9_variance32x64_sse2;
+
+    vp9_variance32x32 = vp9_variance32x32_c;
+    if (flags & HAS_SSE2) vp9_variance32x32 = vp9_variance32x32_sse2;
+
+    vp9_variance64x64 = vp9_variance64x64_c;
+    if (flags & HAS_SSE2) vp9_variance64x64 = vp9_variance64x64_sse2;
+
+    vp9_variance16x16 = vp9_variance16x16_c;
+    if (flags & HAS_MMX) vp9_variance16x16 = vp9_variance16x16_mmx;
+    if (flags & HAS_SSE2) vp9_variance16x16 = vp9_variance16x16_sse2;
+
+    vp9_variance16x8 = vp9_variance16x8_c;
+    if (flags & HAS_MMX) vp9_variance16x8 = vp9_variance16x8_mmx;
+    if (flags & HAS_SSE2) vp9_variance16x8 = vp9_variance16x8_sse2;
+
+    vp9_variance8x16 = vp9_variance8x16_c;
+    if (flags & HAS_MMX) vp9_variance8x16 = vp9_variance8x16_mmx;
+    if (flags & HAS_SSE2) vp9_variance8x16 = vp9_variance8x16_sse2;
+
+    vp9_variance8x8 = vp9_variance8x8_c;
+    if (flags & HAS_MMX) vp9_variance8x8 = vp9_variance8x8_mmx;
+    if (flags & HAS_SSE2) vp9_variance8x8 = vp9_variance8x8_sse2;
+
+    vp9_get_sse_sum_8x8 = vp9_get_sse_sum_8x8_c;
+    if (flags & HAS_SSE2) vp9_get_sse_sum_8x8 = vp9_get8x8var_sse2;
+
+    vp9_variance8x4 = vp9_variance8x4_c;
+    if (flags & HAS_SSE2) vp9_variance8x4 = vp9_variance8x4_sse2;
+
+    vp9_variance4x8 = vp9_variance4x8_c;
+    if (flags & HAS_SSE2) vp9_variance4x8 = vp9_variance4x8_sse2;
+
+    vp9_variance4x4 = vp9_variance4x4_c;
+    if (flags & HAS_MMX) vp9_variance4x4 = vp9_variance4x4_mmx;
+    if (flags & HAS_SSE2) vp9_variance4x4 = vp9_variance4x4_sse2;
+
+    vp9_sub_pixel_variance64x64 = vp9_sub_pixel_variance64x64_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_variance64x64 = vp9_sub_pixel_variance64x64_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance64x64 = vp9_sub_pixel_variance64x64_ssse3;
+
+    vp9_sub_pixel_avg_variance64x64 = vp9_sub_pixel_avg_variance64x64_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_avg_variance64x64 = vp9_sub_pixel_avg_variance64x64_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance64x64 = vp9_sub_pixel_avg_variance64x64_ssse3;
+
+    vp9_sub_pixel_variance32x64 = vp9_sub_pixel_variance32x64_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_variance32x64 = vp9_sub_pixel_variance32x64_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance32x64 = vp9_sub_pixel_variance32x64_ssse3;
+
+    vp9_sub_pixel_avg_variance32x64 = vp9_sub_pixel_avg_variance32x64_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_avg_variance32x64 = vp9_sub_pixel_avg_variance32x64_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance32x64 = vp9_sub_pixel_avg_variance32x64_ssse3;
+
+    vp9_sub_pixel_variance64x32 = vp9_sub_pixel_variance64x32_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_variance64x32 = vp9_sub_pixel_variance64x32_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance64x32 = vp9_sub_pixel_variance64x32_ssse3;
+
+    vp9_sub_pixel_avg_variance64x32 = vp9_sub_pixel_avg_variance64x32_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_avg_variance64x32 = vp9_sub_pixel_avg_variance64x32_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance64x32 = vp9_sub_pixel_avg_variance64x32_ssse3;
+
+    vp9_sub_pixel_variance32x16 = vp9_sub_pixel_variance32x16_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_variance32x16 = vp9_sub_pixel_variance32x16_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance32x16 = vp9_sub_pixel_variance32x16_ssse3;
+
+    vp9_sub_pixel_avg_variance32x16 = vp9_sub_pixel_avg_variance32x16_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_avg_variance32x16 = vp9_sub_pixel_avg_variance32x16_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance32x16 = vp9_sub_pixel_avg_variance32x16_ssse3;
+
+    vp9_sub_pixel_variance16x32 = vp9_sub_pixel_variance16x32_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_variance16x32 = vp9_sub_pixel_variance16x32_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance16x32 = vp9_sub_pixel_variance16x32_ssse3;
+
+    vp9_sub_pixel_avg_variance16x32 = vp9_sub_pixel_avg_variance16x32_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_avg_variance16x32 = vp9_sub_pixel_avg_variance16x32_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance16x32 = vp9_sub_pixel_avg_variance16x32_ssse3;
+
+    vp9_sub_pixel_variance32x32 = vp9_sub_pixel_variance32x32_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_variance32x32 = vp9_sub_pixel_variance32x32_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance32x32 = vp9_sub_pixel_variance32x32_ssse3;
+
+    vp9_sub_pixel_avg_variance32x32 = vp9_sub_pixel_avg_variance32x32_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_avg_variance32x32 = vp9_sub_pixel_avg_variance32x32_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance32x32 = vp9_sub_pixel_avg_variance32x32_ssse3;
+
+    vp9_sub_pixel_variance16x16 = vp9_sub_pixel_variance16x16_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_variance16x16 = vp9_sub_pixel_variance16x16_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance16x16 = vp9_sub_pixel_variance16x16_ssse3;
+
+    vp9_sub_pixel_avg_variance16x16 = vp9_sub_pixel_avg_variance16x16_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_avg_variance16x16 = vp9_sub_pixel_avg_variance16x16_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance16x16 = vp9_sub_pixel_avg_variance16x16_ssse3;
+
+    vp9_sub_pixel_variance8x16 = vp9_sub_pixel_variance8x16_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_variance8x16 = vp9_sub_pixel_variance8x16_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance8x16 = vp9_sub_pixel_variance8x16_ssse3;
+
+    vp9_sub_pixel_avg_variance8x16 = vp9_sub_pixel_avg_variance8x16_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_avg_variance8x16 = vp9_sub_pixel_avg_variance8x16_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance8x16 = vp9_sub_pixel_avg_variance8x16_ssse3;
+
+    vp9_sub_pixel_variance16x8 = vp9_sub_pixel_variance16x8_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_variance16x8 = vp9_sub_pixel_variance16x8_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance16x8 = vp9_sub_pixel_variance16x8_ssse3;
+
+    vp9_sub_pixel_avg_variance16x8 = vp9_sub_pixel_avg_variance16x8_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_avg_variance16x8 = vp9_sub_pixel_avg_variance16x8_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance16x8 = vp9_sub_pixel_avg_variance16x8_ssse3;
+
+    vp9_sub_pixel_variance8x8 = vp9_sub_pixel_variance8x8_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_variance8x8 = vp9_sub_pixel_variance8x8_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance8x8 = vp9_sub_pixel_variance8x8_ssse3;
+
+    vp9_sub_pixel_avg_variance8x8 = vp9_sub_pixel_avg_variance8x8_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_avg_variance8x8 = vp9_sub_pixel_avg_variance8x8_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance8x8 = vp9_sub_pixel_avg_variance8x8_ssse3;
+
+    vp9_sub_pixel_variance8x4 = vp9_sub_pixel_variance8x4_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_variance8x4 = vp9_sub_pixel_variance8x4_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance8x4 = vp9_sub_pixel_variance8x4_ssse3;
+
+    vp9_sub_pixel_avg_variance8x4 = vp9_sub_pixel_avg_variance8x4_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_avg_variance8x4 = vp9_sub_pixel_avg_variance8x4_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance8x4 = vp9_sub_pixel_avg_variance8x4_ssse3;
+
+    vp9_sub_pixel_variance4x8 = vp9_sub_pixel_variance4x8_c;
+    if (flags & HAS_SSE) vp9_sub_pixel_variance4x8 = vp9_sub_pixel_variance4x8_sse;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance4x8 = vp9_sub_pixel_variance4x8_ssse3;
+
+    vp9_sub_pixel_avg_variance4x8 = vp9_sub_pixel_avg_variance4x8_c;
+    if (flags & HAS_SSE) vp9_sub_pixel_avg_variance4x8 = vp9_sub_pixel_avg_variance4x8_sse;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance4x8 = vp9_sub_pixel_avg_variance4x8_ssse3;
+
+    vp9_sub_pixel_variance4x4 = vp9_sub_pixel_variance4x4_c;
+    if (flags & HAS_SSE) vp9_sub_pixel_variance4x4 = vp9_sub_pixel_variance4x4_sse;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance4x4 = vp9_sub_pixel_variance4x4_ssse3;
+
+    vp9_sub_pixel_avg_variance4x4 = vp9_sub_pixel_avg_variance4x4_c;
+    if (flags & HAS_SSE) vp9_sub_pixel_avg_variance4x4 = vp9_sub_pixel_avg_variance4x4_sse;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance4x4 = vp9_sub_pixel_avg_variance4x4_ssse3;
+
+    vp9_sad64x64 = vp9_sad64x64_c;
+    if (flags & HAS_SSE2) vp9_sad64x64 = vp9_sad64x64_sse2;
+
+    vp9_sad32x64 = vp9_sad32x64_c;
+    if (flags & HAS_SSE2) vp9_sad32x64 = vp9_sad32x64_sse2;
+
+    vp9_sad64x32 = vp9_sad64x32_c;
+    if (flags & HAS_SSE2) vp9_sad64x32 = vp9_sad64x32_sse2;
+
+    vp9_sad32x16 = vp9_sad32x16_c;
+    if (flags & HAS_SSE2) vp9_sad32x16 = vp9_sad32x16_sse2;
+
+    vp9_sad16x32 = vp9_sad16x32_c;
+    if (flags & HAS_SSE2) vp9_sad16x32 = vp9_sad16x32_sse2;
+
+    vp9_sad32x32 = vp9_sad32x32_c;
+    if (flags & HAS_SSE2) vp9_sad32x32 = vp9_sad32x32_sse2;
+
+    vp9_sad16x16 = vp9_sad16x16_c;
+    if (flags & HAS_MMX) vp9_sad16x16 = vp9_sad16x16_mmx;
+    if (flags & HAS_SSE2) vp9_sad16x16 = vp9_sad16x16_sse2;
+
+    vp9_sad16x8 = vp9_sad16x8_c;
+    if (flags & HAS_MMX) vp9_sad16x8 = vp9_sad16x8_mmx;
+    if (flags & HAS_SSE2) vp9_sad16x8 = vp9_sad16x8_sse2;
+
+    vp9_sad8x16 = vp9_sad8x16_c;
+    if (flags & HAS_MMX) vp9_sad8x16 = vp9_sad8x16_mmx;
+    if (flags & HAS_SSE2) vp9_sad8x16 = vp9_sad8x16_sse2;
+
+    vp9_sad8x8 = vp9_sad8x8_c;
+    if (flags & HAS_MMX) vp9_sad8x8 = vp9_sad8x8_mmx;
+    if (flags & HAS_SSE2) vp9_sad8x8 = vp9_sad8x8_sse2;
+
+    vp9_sad8x4 = vp9_sad8x4_c;
+    if (flags & HAS_SSE2) vp9_sad8x4 = vp9_sad8x4_sse2;
+
+    vp9_sad4x8 = vp9_sad4x8_c;
+    if (flags & HAS_SSE) vp9_sad4x8 = vp9_sad4x8_sse;
+
+    vp9_sad4x4 = vp9_sad4x4_c;
+    if (flags & HAS_MMX) vp9_sad4x4 = vp9_sad4x4_mmx;
+    if (flags & HAS_SSE) vp9_sad4x4 = vp9_sad4x4_sse;
+
+    vp9_sad64x64_avg = vp9_sad64x64_avg_c;
+    if (flags & HAS_SSE2) vp9_sad64x64_avg = vp9_sad64x64_avg_sse2;
+
+    vp9_sad32x64_avg = vp9_sad32x64_avg_c;
+    if (flags & HAS_SSE2) vp9_sad32x64_avg = vp9_sad32x64_avg_sse2;
+
+    vp9_sad64x32_avg = vp9_sad64x32_avg_c;
+    if (flags & HAS_SSE2) vp9_sad64x32_avg = vp9_sad64x32_avg_sse2;
+
+    vp9_sad32x16_avg = vp9_sad32x16_avg_c;
+    if (flags & HAS_SSE2) vp9_sad32x16_avg = vp9_sad32x16_avg_sse2;
+
+    vp9_sad16x32_avg = vp9_sad16x32_avg_c;
+    if (flags & HAS_SSE2) vp9_sad16x32_avg = vp9_sad16x32_avg_sse2;
+
+    vp9_sad32x32_avg = vp9_sad32x32_avg_c;
+    if (flags & HAS_SSE2) vp9_sad32x32_avg = vp9_sad32x32_avg_sse2;
+
+    vp9_sad16x16_avg = vp9_sad16x16_avg_c;
+    if (flags & HAS_SSE2) vp9_sad16x16_avg = vp9_sad16x16_avg_sse2;
+
+    vp9_sad16x8_avg = vp9_sad16x8_avg_c;
+    if (flags & HAS_SSE2) vp9_sad16x8_avg = vp9_sad16x8_avg_sse2;
+
+    vp9_sad8x16_avg = vp9_sad8x16_avg_c;
+    if (flags & HAS_SSE2) vp9_sad8x16_avg = vp9_sad8x16_avg_sse2;
+
+    vp9_sad8x8_avg = vp9_sad8x8_avg_c;
+    if (flags & HAS_SSE2) vp9_sad8x8_avg = vp9_sad8x8_avg_sse2;
+
+    vp9_sad8x4_avg = vp9_sad8x4_avg_c;
+    if (flags & HAS_SSE2) vp9_sad8x4_avg = vp9_sad8x4_avg_sse2;
+
+    vp9_sad4x8_avg = vp9_sad4x8_avg_c;
+    if (flags & HAS_SSE) vp9_sad4x8_avg = vp9_sad4x8_avg_sse;
+
+    vp9_sad4x4_avg = vp9_sad4x4_avg_c;
+    if (flags & HAS_SSE) vp9_sad4x4_avg = vp9_sad4x4_avg_sse;
+
+    vp9_variance_halfpixvar16x16_h = vp9_variance_halfpixvar16x16_h_c;
+    if (flags & HAS_SSE2) vp9_variance_halfpixvar16x16_h = vp9_variance_halfpixvar16x16_h_sse2;
+
+    vp9_variance_halfpixvar16x16_v = vp9_variance_halfpixvar16x16_v_c;
+    if (flags & HAS_SSE2) vp9_variance_halfpixvar16x16_v = vp9_variance_halfpixvar16x16_v_sse2;
+
+    vp9_variance_halfpixvar16x16_hv = vp9_variance_halfpixvar16x16_hv_c;
+    if (flags & HAS_SSE2) vp9_variance_halfpixvar16x16_hv = vp9_variance_halfpixvar16x16_hv_sse2;
+
+
+
+
+
+
+
+
+
+    vp9_sad16x16x3 = vp9_sad16x16x3_c;
+    if (flags & HAS_SSE3) vp9_sad16x16x3 = vp9_sad16x16x3_sse3;
+    if (flags & HAS_SSSE3) vp9_sad16x16x3 = vp9_sad16x16x3_ssse3;
+
+    vp9_sad16x8x3 = vp9_sad16x8x3_c;
+    if (flags & HAS_SSE3) vp9_sad16x8x3 = vp9_sad16x8x3_sse3;
+    if (flags & HAS_SSSE3) vp9_sad16x8x3 = vp9_sad16x8x3_ssse3;
+
+    vp9_sad8x16x3 = vp9_sad8x16x3_c;
+    if (flags & HAS_SSE3) vp9_sad8x16x3 = vp9_sad8x16x3_sse3;
+
+    vp9_sad8x8x3 = vp9_sad8x8x3_c;
+    if (flags & HAS_SSE3) vp9_sad8x8x3 = vp9_sad8x8x3_sse3;
+
+    vp9_sad4x4x3 = vp9_sad4x4x3_c;
+    if (flags & HAS_SSE3) vp9_sad4x4x3 = vp9_sad4x4x3_sse3;
+
+
+
+
+
+
+
+
+
+
+    vp9_sad64x64x4d = vp9_sad64x64x4d_c;
+    if (flags & HAS_SSE2) vp9_sad64x64x4d = vp9_sad64x64x4d_sse2;
+
+    vp9_sad32x64x4d = vp9_sad32x64x4d_c;
+    if (flags & HAS_SSE2) vp9_sad32x64x4d = vp9_sad32x64x4d_sse2;
+
+    vp9_sad64x32x4d = vp9_sad64x32x4d_c;
+    if (flags & HAS_SSE2) vp9_sad64x32x4d = vp9_sad64x32x4d_sse2;
+
+    vp9_sad32x16x4d = vp9_sad32x16x4d_c;
+    if (flags & HAS_SSE2) vp9_sad32x16x4d = vp9_sad32x16x4d_sse2;
+
+    vp9_sad16x32x4d = vp9_sad16x32x4d_c;
+    if (flags & HAS_SSE2) vp9_sad16x32x4d = vp9_sad16x32x4d_sse2;
+
+    vp9_sad32x32x4d = vp9_sad32x32x4d_c;
+    if (flags & HAS_SSE2) vp9_sad32x32x4d = vp9_sad32x32x4d_sse2;
+
+    vp9_sad16x16x4d = vp9_sad16x16x4d_c;
+    if (flags & HAS_SSE2) vp9_sad16x16x4d = vp9_sad16x16x4d_sse2;
+
+    vp9_sad16x8x4d = vp9_sad16x8x4d_c;
+    if (flags & HAS_SSE2) vp9_sad16x8x4d = vp9_sad16x8x4d_sse2;
+
+    vp9_sad8x16x4d = vp9_sad8x16x4d_c;
+    if (flags & HAS_SSE2) vp9_sad8x16x4d = vp9_sad8x16x4d_sse2;
+
+    vp9_sad8x8x4d = vp9_sad8x8x4d_c;
+    if (flags & HAS_SSE2) vp9_sad8x8x4d = vp9_sad8x8x4d_sse2;
+
+    vp9_sad8x4x4d = vp9_sad8x4x4d_c;
+    if (flags & HAS_SSE2) vp9_sad8x4x4d = vp9_sad8x4x4d_sse2;
+
+    vp9_sad4x8x4d = vp9_sad4x8x4d_c;
+    if (flags & HAS_SSE) vp9_sad4x8x4d = vp9_sad4x8x4d_sse;
+
+    vp9_sad4x4x4d = vp9_sad4x4x4d_c;
+    if (flags & HAS_SSE) vp9_sad4x4x4d = vp9_sad4x4x4d_sse;
+
+    vp9_mse16x16 = vp9_mse16x16_c;
+    if (flags & HAS_MMX) vp9_mse16x16 = vp9_mse16x16_mmx;
+    if (flags & HAS_SSE2) vp9_mse16x16 = vp9_mse16x16_sse2;
+
+
+
+
+
+
+    vp9_get_mb_ss = vp9_get_mb_ss_c;
+    if (flags & HAS_MMX) vp9_get_mb_ss = vp9_get_mb_ss_mmx;
+    if (flags & HAS_SSE2) vp9_get_mb_ss = vp9_get_mb_ss_sse2;
+
+    vp9_block_error = vp9_block_error_c;
+    if (flags & HAS_SSE2) vp9_block_error = vp9_block_error_sse2;
+
+    vp9_subtract_block = vp9_subtract_block_c;
+    if (flags & HAS_SSE2) vp9_subtract_block = vp9_subtract_block_sse2;
+
+
+
+    vp9_short_fht4x4 = vp9_short_fht4x4_c;
+    if (flags & HAS_SSE2) vp9_short_fht4x4 = vp9_short_fht4x4_sse2;
+
+    vp9_short_fht8x8 = vp9_short_fht8x8_c;
+    if (flags & HAS_SSE2) vp9_short_fht8x8 = vp9_short_fht8x8_sse2;
+
+    vp9_short_fht16x16 = vp9_short_fht16x16_c;
+    if (flags & HAS_SSE2) vp9_short_fht16x16 = vp9_short_fht16x16_sse2;
+
+
+    vp9_fdct4x4 = vp9_fdct4x4_c;
+    if (flags & HAS_SSE2) vp9_fdct4x4 = vp9_fdct4x4_sse2;
+
+    vp9_fdct8x8 = vp9_fdct8x8_c;
+    if (flags & HAS_SSE2) vp9_fdct8x8 = vp9_fdct8x8_sse2;
+
+    vp9_fdct16x16 = vp9_fdct16x16_c;
+    if (flags & HAS_SSE2) vp9_fdct16x16 = vp9_fdct16x16_sse2;
+
+    vp9_fdct32x32 = vp9_fdct32x32_c;
+    if (flags & HAS_SSE2) vp9_fdct32x32 = vp9_fdct32x32_sse2;
+
+    vp9_fdct32x32_rd = vp9_fdct32x32_rd_c;
+    if (flags & HAS_SSE2) vp9_fdct32x32_rd = vp9_fdct32x32_rd_sse2;
+
+    vp9_full_search_sad = vp9_full_search_sad_c;
+    if (flags & HAS_SSE3) vp9_full_search_sad = vp9_full_search_sadx3;
+    if (flags & HAS_SSE4_1) vp9_full_search_sad = vp9_full_search_sadx8;
+
+    vp9_refining_search_sad = vp9_refining_search_sad_c;
+    if (flags & HAS_SSE3) vp9_refining_search_sad = vp9_refining_search_sadx4;
+
+    vp9_diamond_search_sad = vp9_diamond_search_sad_c;
+    if (flags & HAS_SSE3) vp9_diamond_search_sad = vp9_diamond_search_sadx4;
+
+    vp9_temporal_filter_apply = vp9_temporal_filter_apply_c;
+    if (flags & HAS_SSE2) vp9_temporal_filter_apply = vp9_temporal_filter_apply_sse2;
+}
+#endif
+#endif
diff --git a/media/libvpx/vp9_rtcd_x86_64-win64-gcc.h b/media/libvpx/vp9_rtcd_x86_64-win64-gcc.h
new file mode 100644
index 00000000000..db6bbc2485e
--- /dev/null
+++ b/media/libvpx/vp9_rtcd_x86_64-win64-gcc.h
@@ -0,0 +1,1228 @@
+#ifndef VP9_RTCD_H_
+#define VP9_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#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]
+union int_mv;
+struct yv12_buffer_config;
+
+void vp9_d207_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d207_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d207_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d45_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d45_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d45_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d63_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d63_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d63_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_h_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_h_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_d153_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d153_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d153_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_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_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_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_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_d207_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d207_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d207_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d45_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d45_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d45_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d63_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d63_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d63_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_h_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_h_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d153_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d153_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_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_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_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_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_d207_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d207_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d207_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d45_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d45_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d45_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d63_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d63_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d63_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_h_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_h_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_h_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_d153_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d153_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d153_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_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_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_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_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_d207_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d207_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d207_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d45_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d45_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d45_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_d63_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_d63_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_d63_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+void vp9_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_h_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*vp9_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+
+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_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_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_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_tm_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void vp9_tm_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define vp9_tm_predictor_32x32 vp9_tm_predictor_32x32_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_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_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_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_mb_lpf_vertical_edge_w_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+void vp9_mb_lpf_vertical_edge_w_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define vp9_mb_lpf_vertical_edge_w vp9_mb_lpf_vertical_edge_w_sse2
+
+void vp9_mbloop_filter_vertical_edge_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_mbloop_filter_vertical_edge_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+#define vp9_mbloop_filter_vertical_edge vp9_mbloop_filter_vertical_edge_sse2
+
+void vp9_loop_filter_vertical_edge_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_loop_filter_vertical_edge_mmx(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+#define vp9_loop_filter_vertical_edge vp9_loop_filter_vertical_edge_mmx
+
+void vp9_mb_lpf_horizontal_edge_w_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_mb_lpf_horizontal_edge_w_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+#define vp9_mb_lpf_horizontal_edge_w vp9_mb_lpf_horizontal_edge_w_sse2
+
+void vp9_mbloop_filter_horizontal_edge_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_mbloop_filter_horizontal_edge_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+#define vp9_mbloop_filter_horizontal_edge vp9_mbloop_filter_horizontal_edge_sse2
+
+void vp9_loop_filter_horizontal_edge_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+void vp9_loop_filter_horizontal_edge_mmx(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count);
+#define vp9_loop_filter_horizontal_edge vp9_loop_filter_horizontal_edge_mmx
+
+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_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_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_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_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);
+void vp9_convolve8_ssse3(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 (*vp9_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 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);
+void vp9_convolve8_horiz_ssse3(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 (*vp9_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 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);
+void vp9_convolve8_vert_ssse3(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 (*vp9_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 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);
+void vp9_convolve8_avg_ssse3(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 (*vp9_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 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);
+void vp9_convolve8_avg_horiz_ssse3(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 (*vp9_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 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);
+void vp9_convolve8_avg_vert_ssse3(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 (*vp9_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 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_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_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_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_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_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_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_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_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_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_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_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_iwht4x4_16_add_c(const int16_t *input, uint8_t *dest, int dest_stride);
+#define vp9_iwht4x4_16_add vp9_iwht4x4_16_add_c
+
+unsigned int vp9_variance32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance32x16 vp9_variance32x16_sse2
+
+unsigned int vp9_variance16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance16x32 vp9_variance16x32_sse2
+
+unsigned int vp9_variance64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance64x32 vp9_variance64x32_sse2
+
+unsigned int vp9_variance32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance32x64 vp9_variance32x64_sse2
+
+unsigned int vp9_variance32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance32x32 vp9_variance32x32_sse2
+
+unsigned int vp9_variance64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance64x64 vp9_variance64x64_sse2
+
+unsigned int vp9_variance16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance16x16 vp9_variance16x16_sse2
+
+unsigned int vp9_variance16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x8_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance16x8 vp9_variance16x8_sse2
+
+unsigned int vp9_variance8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance8x16 vp9_variance8x16_sse2
+
+unsigned int vp9_variance8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x8_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance8x8 vp9_variance8x8_sse2
+
+void vp9_get_sse_sum_8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+void vp9_get8x8var_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, int *sum);
+#define vp9_get_sse_sum_8x8 vp9_get8x8var_sse2
+
+unsigned int vp9_variance8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance8x4 vp9_variance8x4_sse2
+
+unsigned int vp9_variance4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance4x8 vp9_variance4x8_sse2
+
+unsigned int vp9_variance4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance4x4_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance4x4 vp9_variance4x4_sse2
+
+unsigned int vp9_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance4x8_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance4x8_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance4x4_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+
+unsigned int vp9_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, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance4x4_sse(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+unsigned int vp9_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse, const uint8_t *second_pred);
+
+unsigned int vp9_sad64x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad64x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+#define vp9_sad64x64 vp9_sad64x64_sse2
+
+unsigned int vp9_sad32x64_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x64_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+#define vp9_sad32x64 vp9_sad32x64_sse2
+
+unsigned int vp9_sad64x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad64x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+#define vp9_sad64x32 vp9_sad64x32_sse2
+
+unsigned int vp9_sad32x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+#define vp9_sad32x16 vp9_sad32x16_sse2
+
+unsigned int vp9_sad16x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+#define vp9_sad16x32 vp9_sad16x32_sse2
+
+unsigned int vp9_sad32x32_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad32x32_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+#define vp9_sad32x32 vp9_sad32x32_sse2
+
+unsigned int vp9_sad16x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+#define vp9_sad16x16 vp9_sad16x16_sse2
+
+unsigned int vp9_sad16x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x8_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad16x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+#define vp9_sad16x8 vp9_sad16x8_sse2
+
+unsigned int vp9_sad8x16_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x16_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x16_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+#define vp9_sad8x16 vp9_sad8x16_sse2
+
+unsigned int vp9_sad8x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x8_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x8_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+#define vp9_sad8x8 vp9_sad8x8_sse2
+
+unsigned int vp9_sad8x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad8x4_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+#define vp9_sad8x4 vp9_sad8x4_sse2
+
+unsigned int vp9_sad4x8_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x8_sse(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int max_sad);
+#define vp9_sad4x8 vp9_sad4x8_sse
+
+unsigned int vp9_sad4x4_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x4_mmx(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+unsigned int vp9_sad4x4_sse(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int max_sad);
+#define vp9_sad4x4 vp9_sad4x4_sse
+
+unsigned int vp9_sad64x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad64x64_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+#define vp9_sad64x64_avg vp9_sad64x64_avg_sse2
+
+unsigned int vp9_sad32x64_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x64_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+#define vp9_sad32x64_avg vp9_sad32x64_avg_sse2
+
+unsigned int vp9_sad64x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad64x32_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+#define vp9_sad64x32_avg vp9_sad64x32_avg_sse2
+
+unsigned int vp9_sad32x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x16_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+#define vp9_sad32x16_avg vp9_sad32x16_avg_sse2
+
+unsigned int vp9_sad16x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x32_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+#define vp9_sad16x32_avg vp9_sad16x32_avg_sse2
+
+unsigned int vp9_sad32x32_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad32x32_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+#define vp9_sad32x32_avg vp9_sad32x32_avg_sse2
+
+unsigned int vp9_sad16x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x16_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+#define vp9_sad16x16_avg vp9_sad16x16_avg_sse2
+
+unsigned int vp9_sad16x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad16x8_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+#define vp9_sad16x8_avg vp9_sad16x8_avg_sse2
+
+unsigned int vp9_sad8x16_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x16_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+#define vp9_sad8x16_avg vp9_sad8x16_avg_sse2
+
+unsigned int vp9_sad8x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x8_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+#define vp9_sad8x8_avg vp9_sad8x8_avg_sse2
+
+unsigned int vp9_sad8x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad8x4_avg_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+#define vp9_sad8x4_avg vp9_sad8x4_avg_sse2
+
+unsigned int vp9_sad4x8_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad4x8_avg_sse(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+#define vp9_sad4x8_avg vp9_sad4x8_avg_sse
+
+unsigned int vp9_sad4x4_avg_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+unsigned int vp9_sad4x4_avg_sse(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, const uint8_t *second_pred, unsigned int max_sad);
+#define vp9_sad4x4_avg vp9_sad4x4_avg_sse
+
+unsigned int vp9_variance_halfpixvar16x16_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance_halfpixvar16x16_h_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar16x16_h vp9_variance_halfpixvar16x16_h_sse2
+
+unsigned int vp9_variance_halfpixvar16x16_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance_halfpixvar16x16_v_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar16x16_v vp9_variance_halfpixvar16x16_v_sse2
+
+unsigned int vp9_variance_halfpixvar16x16_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+unsigned int vp9_variance_halfpixvar16x16_hv_sse2(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar16x16_hv vp9_variance_halfpixvar16x16_hv_sse2
+
+unsigned int vp9_variance_halfpixvar64x64_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar64x64_h vp9_variance_halfpixvar64x64_h_c
+
+unsigned int vp9_variance_halfpixvar64x64_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar64x64_v vp9_variance_halfpixvar64x64_v_c
+
+unsigned int vp9_variance_halfpixvar64x64_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar64x64_hv vp9_variance_halfpixvar64x64_hv_c
+
+unsigned int vp9_variance_halfpixvar32x32_h_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar32x32_h vp9_variance_halfpixvar32x32_h_c
+
+unsigned int vp9_variance_halfpixvar32x32_v_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar32x32_v vp9_variance_halfpixvar32x32_v_c
+
+unsigned int vp9_variance_halfpixvar32x32_hv_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_variance_halfpixvar32x32_hv vp9_variance_halfpixvar32x32_hv_c
+
+void vp9_sad64x64x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp9_sad64x64x3 vp9_sad64x64x3_c
+
+void vp9_sad32x32x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+#define vp9_sad32x32x3 vp9_sad32x32x3_c
+
+void vp9_sad16x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x16x3_sse3(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x16x3_ssse3(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad16x16x3)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad16x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x8x3_sse3(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x8x3_ssse3(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad16x8x3)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad8x16x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp9_sad8x16x3_sse3(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad8x16x3)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad8x8x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp9_sad8x8x3_sse3(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad8x8x3)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad4x4x3_c(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+void vp9_sad4x4x3_sse3(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+RTCD_EXTERN void (*vp9_sad4x4x3)(const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr, int  ref_stride, unsigned int *sad_array);
+
+void vp9_sad64x64x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad64x64x8 vp9_sad64x64x8_c
+
+void vp9_sad32x32x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad32x32x8 vp9_sad32x32x8_c
+
+void vp9_sad16x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad16x16x8 vp9_sad16x16x8_c
+
+void vp9_sad16x8x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad16x8x8 vp9_sad16x8x8_c
+
+void vp9_sad8x16x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad8x16x8 vp9_sad8x16x8_c
+
+void vp9_sad8x8x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad8x8x8 vp9_sad8x8x8_c
+
+void vp9_sad8x4x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vp9_sad8x4x8 vp9_sad8x4x8_c
+
+void vp9_sad4x8x8_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, uint32_t *sad_array);
+#define vp9_sad4x8x8 vp9_sad4x8x8_c
+
+void vp9_sad4x4x8_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int  ref_stride, uint32_t *sad_array);
+#define vp9_sad4x4x8 vp9_sad4x4x8_c
+
+void vp9_sad64x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad64x64x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad64x64x4d vp9_sad64x64x4d_sse2
+
+void vp9_sad32x64x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad32x64x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad32x64x4d vp9_sad32x64x4d_sse2
+
+void vp9_sad64x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad64x32x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad64x32x4d vp9_sad64x32x4d_sse2
+
+void vp9_sad32x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad32x16x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad32x16x4d vp9_sad32x16x4d_sse2
+
+void vp9_sad16x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x32x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad16x32x4d vp9_sad16x32x4d_sse2
+
+void vp9_sad32x32x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad32x32x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad32x32x4d vp9_sad32x32x4d_sse2
+
+void vp9_sad16x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x16x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad16x16x4d vp9_sad16x16x4d_sse2
+
+void vp9_sad16x8x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad16x8x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad16x8x4d vp9_sad16x8x4d_sse2
+
+void vp9_sad8x16x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad8x16x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad8x16x4d vp9_sad8x16x4d_sse2
+
+void vp9_sad8x8x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad8x8x4d_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad8x8x4d vp9_sad8x8x4d_sse2
+
+void vp9_sad8x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
+void vp9_sad8x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
+#define vp9_sad8x4x4d vp9_sad8x4x4d_sse2
+
+void vp9_sad4x8x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
+void vp9_sad4x8x4d_sse(const uint8_t *src_ptr, int src_stride, const uint8_t* const ref_ptr[], int ref_stride, unsigned int *sad_array);
+#define vp9_sad4x8x4d vp9_sad4x8x4d_sse
+
+void vp9_sad4x4x4d_c(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+void vp9_sad4x4x4d_sse(const uint8_t *src_ptr, int  src_stride, const uint8_t* const ref_ptr[], int  ref_stride, unsigned int *sad_array);
+#define vp9_sad4x4x4d vp9_sad4x4x4d_sse
+
+unsigned int vp9_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vp9_mse16x16_mmx(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int vp9_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vp9_mse16x16 vp9_mse16x16_sse2
+
+unsigned int vp9_mse8x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vp9_mse8x16 vp9_mse8x16_c
+
+unsigned int vp9_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vp9_mse16x8 vp9_mse16x8_c
+
+unsigned int vp9_mse8x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+#define vp9_mse8x8 vp9_mse8x8_c
+
+unsigned int vp9_sub_pixel_mse64x64_c(const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_mse64x64 vp9_sub_pixel_mse64x64_c
+
+unsigned int vp9_sub_pixel_mse32x32_c(const uint8_t *src_ptr, int  source_stride, int  xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, unsigned int *sse);
+#define vp9_sub_pixel_mse32x32 vp9_sub_pixel_mse32x32_c
+
+unsigned int vp9_get_mb_ss_c(const int16_t *);
+unsigned int vp9_get_mb_ss_mmx(const int16_t *);
+unsigned int vp9_get_mb_ss_sse2(const int16_t *);
+#define vp9_get_mb_ss vp9_get_mb_ss_sse2
+
+int64_t vp9_block_error_c(int16_t *coeff, int16_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+int64_t vp9_block_error_sse2(int16_t *coeff, int16_t *dqcoeff, intptr_t block_size, int64_t *ssz);
+#define vp9_block_error vp9_block_error_sse2
+
+void vp9_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);
+void vp9_subtract_block_sse2(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 vp9_subtract_block vp9_subtract_block_sse2
+
+void vp9_quantize_b_c(const int16_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, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_quantize_b_ssse3(const int16_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, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*vp9_quantize_b)(const int16_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, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void vp9_quantize_b_32x32_c(const int16_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, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+void vp9_quantize_b_32x32_ssse3(const int16_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, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+RTCD_EXTERN void (*vp9_quantize_b_32x32)(const int16_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, int16_t *qcoeff_ptr, int16_t *dqcoeff_ptr, const int16_t *dequant_ptr, int zbin_oq_value, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
+
+void vp9_short_fht4x4_c(const int16_t *input, int16_t *output, int stride, int tx_type);
+void vp9_short_fht4x4_sse2(const int16_t *input, int16_t *output, int stride, int tx_type);
+#define vp9_short_fht4x4 vp9_short_fht4x4_sse2
+
+void vp9_short_fht8x8_c(const int16_t *input, int16_t *output, int stride, int tx_type);
+void vp9_short_fht8x8_sse2(const int16_t *input, int16_t *output, int stride, int tx_type);
+#define vp9_short_fht8x8 vp9_short_fht8x8_sse2
+
+void vp9_short_fht16x16_c(const int16_t *input, int16_t *output, int stride, int tx_type);
+void vp9_short_fht16x16_sse2(const int16_t *input, int16_t *output, int stride, int tx_type);
+#define vp9_short_fht16x16 vp9_short_fht16x16_sse2
+
+void vp9_fwht4x4_c(const int16_t *input, int16_t *output, int stride);
+#define vp9_fwht4x4 vp9_fwht4x4_c
+
+void vp9_fdct4x4_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct4x4_sse2(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct4x4 vp9_fdct4x4_sse2
+
+void vp9_fdct8x8_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct8x8_sse2(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct8x8 vp9_fdct8x8_sse2
+
+void vp9_fdct16x16_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct16x16_sse2(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct16x16 vp9_fdct16x16_sse2
+
+void vp9_fdct32x32_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct32x32_sse2(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct32x32 vp9_fdct32x32_sse2
+
+void vp9_fdct32x32_rd_c(const int16_t *input, int16_t *output, int stride);
+void vp9_fdct32x32_rd_sse2(const int16_t *input, int16_t *output, int stride);
+#define vp9_fdct32x32_rd vp9_fdct32x32_rd_sse2
+
+int vp9_full_search_sad_c(struct macroblock *x, union int_mv *ref_mv, int sad_per_bit, int distance, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv, int n);
+int vp9_full_search_sadx3(struct macroblock *x, union int_mv *ref_mv, int sad_per_bit, int distance, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv, int n);
+int vp9_full_search_sadx8(struct macroblock *x, union int_mv *ref_mv, int sad_per_bit, int distance, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv, int n);
+RTCD_EXTERN int (*vp9_full_search_sad)(struct macroblock *x, union int_mv *ref_mv, int sad_per_bit, int distance, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv, int n);
+
+int vp9_refining_search_sad_c(struct macroblock *x, union int_mv *ref_mv, int sad_per_bit, int distance, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv);
+int vp9_refining_search_sadx4(struct macroblock *x, union int_mv *ref_mv, int sad_per_bit, int distance, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv);
+RTCD_EXTERN int (*vp9_refining_search_sad)(struct macroblock *x, union int_mv *ref_mv, int sad_per_bit, int distance, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv);
+
+int vp9_diamond_search_sad_c(struct macroblock *x, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv);
+int vp9_diamond_search_sadx4(struct macroblock *x, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv);
+RTCD_EXTERN int (*vp9_diamond_search_sad)(struct macroblock *x, union int_mv *ref_mv, union int_mv *best_mv, int search_param, int sad_per_bit, int *num00, struct vp9_variance_vtable *fn_ptr, DEC_MVCOSTS, union int_mv *center_mv);
+
+void vp9_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+void vp9_temporal_filter_apply_sse2(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_size, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
+#define vp9_temporal_filter_apply vp9_temporal_filter_apply_sse2
+
+void vp9_yv12_copy_partial_frame_c(struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int fraction);
+#define vp9_yv12_copy_partial_frame vp9_yv12_copy_partial_frame_c
+
+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;
+
+    vp9_d207_predictor_4x4 = vp9_d207_predictor_4x4_c;
+    if (flags & HAS_SSSE3) vp9_d207_predictor_4x4 = vp9_d207_predictor_4x4_ssse3;
+
+    vp9_d45_predictor_4x4 = vp9_d45_predictor_4x4_c;
+    if (flags & HAS_SSSE3) vp9_d45_predictor_4x4 = vp9_d45_predictor_4x4_ssse3;
+
+    vp9_d63_predictor_4x4 = vp9_d63_predictor_4x4_c;
+    if (flags & HAS_SSSE3) vp9_d63_predictor_4x4 = vp9_d63_predictor_4x4_ssse3;
+
+    vp9_h_predictor_4x4 = vp9_h_predictor_4x4_c;
+    if (flags & HAS_SSSE3) vp9_h_predictor_4x4 = vp9_h_predictor_4x4_ssse3;
+
+
+
+    vp9_d153_predictor_4x4 = vp9_d153_predictor_4x4_c;
+    if (flags & HAS_SSSE3) vp9_d153_predictor_4x4 = vp9_d153_predictor_4x4_ssse3;
+
+
+
+
+
+
+
+    vp9_d207_predictor_8x8 = vp9_d207_predictor_8x8_c;
+    if (flags & HAS_SSSE3) vp9_d207_predictor_8x8 = vp9_d207_predictor_8x8_ssse3;
+
+    vp9_d45_predictor_8x8 = vp9_d45_predictor_8x8_c;
+    if (flags & HAS_SSSE3) vp9_d45_predictor_8x8 = vp9_d45_predictor_8x8_ssse3;
+
+    vp9_d63_predictor_8x8 = vp9_d63_predictor_8x8_c;
+    if (flags & HAS_SSSE3) vp9_d63_predictor_8x8 = vp9_d63_predictor_8x8_ssse3;
+
+    vp9_h_predictor_8x8 = vp9_h_predictor_8x8_c;
+    if (flags & HAS_SSSE3) vp9_h_predictor_8x8 = vp9_h_predictor_8x8_ssse3;
+
+
+
+    vp9_d153_predictor_8x8 = vp9_d153_predictor_8x8_c;
+    if (flags & HAS_SSSE3) vp9_d153_predictor_8x8 = vp9_d153_predictor_8x8_ssse3;
+
+
+
+
+
+
+
+    vp9_d207_predictor_16x16 = vp9_d207_predictor_16x16_c;
+    if (flags & HAS_SSSE3) vp9_d207_predictor_16x16 = vp9_d207_predictor_16x16_ssse3;
+
+    vp9_d45_predictor_16x16 = vp9_d45_predictor_16x16_c;
+    if (flags & HAS_SSSE3) vp9_d45_predictor_16x16 = vp9_d45_predictor_16x16_ssse3;
+
+    vp9_d63_predictor_16x16 = vp9_d63_predictor_16x16_c;
+    if (flags & HAS_SSSE3) vp9_d63_predictor_16x16 = vp9_d63_predictor_16x16_ssse3;
+
+    vp9_h_predictor_16x16 = vp9_h_predictor_16x16_c;
+    if (flags & HAS_SSSE3) vp9_h_predictor_16x16 = vp9_h_predictor_16x16_ssse3;
+
+
+
+    vp9_d153_predictor_16x16 = vp9_d153_predictor_16x16_c;
+    if (flags & HAS_SSSE3) vp9_d153_predictor_16x16 = vp9_d153_predictor_16x16_ssse3;
+
+
+
+
+
+
+
+    vp9_d207_predictor_32x32 = vp9_d207_predictor_32x32_c;
+    if (flags & HAS_SSSE3) vp9_d207_predictor_32x32 = vp9_d207_predictor_32x32_ssse3;
+
+    vp9_d45_predictor_32x32 = vp9_d45_predictor_32x32_c;
+    if (flags & HAS_SSSE3) vp9_d45_predictor_32x32 = vp9_d45_predictor_32x32_ssse3;
+
+    vp9_d63_predictor_32x32 = vp9_d63_predictor_32x32_c;
+    if (flags & HAS_SSSE3) vp9_d63_predictor_32x32 = vp9_d63_predictor_32x32_ssse3;
+
+    vp9_h_predictor_32x32 = vp9_h_predictor_32x32_c;
+    if (flags & HAS_SSSE3) vp9_h_predictor_32x32 = vp9_h_predictor_32x32_ssse3;
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+    vp9_convolve8 = vp9_convolve8_sse2;
+    if (flags & HAS_SSSE3) vp9_convolve8 = vp9_convolve8_ssse3;
+
+    vp9_convolve8_horiz = vp9_convolve8_horiz_sse2;
+    if (flags & HAS_SSSE3) vp9_convolve8_horiz = vp9_convolve8_horiz_ssse3;
+
+    vp9_convolve8_vert = vp9_convolve8_vert_sse2;
+    if (flags & HAS_SSSE3) vp9_convolve8_vert = vp9_convolve8_vert_ssse3;
+
+    vp9_convolve8_avg = vp9_convolve8_avg_sse2;
+    if (flags & HAS_SSSE3) vp9_convolve8_avg = vp9_convolve8_avg_ssse3;
+
+    vp9_convolve8_avg_horiz = vp9_convolve8_avg_horiz_sse2;
+    if (flags & HAS_SSSE3) vp9_convolve8_avg_horiz = vp9_convolve8_avg_horiz_ssse3;
+
+    vp9_convolve8_avg_vert = vp9_convolve8_avg_vert_sse2;
+    if (flags & HAS_SSSE3) vp9_convolve8_avg_vert = vp9_convolve8_avg_vert_ssse3;
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+    vp9_sub_pixel_variance64x64 = vp9_sub_pixel_variance64x64_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance64x64 = vp9_sub_pixel_variance64x64_ssse3;
+
+    vp9_sub_pixel_avg_variance64x64 = vp9_sub_pixel_avg_variance64x64_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance64x64 = vp9_sub_pixel_avg_variance64x64_ssse3;
+
+    vp9_sub_pixel_variance32x64 = vp9_sub_pixel_variance32x64_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance32x64 = vp9_sub_pixel_variance32x64_ssse3;
+
+    vp9_sub_pixel_avg_variance32x64 = vp9_sub_pixel_avg_variance32x64_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance32x64 = vp9_sub_pixel_avg_variance32x64_ssse3;
+
+    vp9_sub_pixel_variance64x32 = vp9_sub_pixel_variance64x32_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance64x32 = vp9_sub_pixel_variance64x32_ssse3;
+
+    vp9_sub_pixel_avg_variance64x32 = vp9_sub_pixel_avg_variance64x32_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance64x32 = vp9_sub_pixel_avg_variance64x32_ssse3;
+
+    vp9_sub_pixel_variance32x16 = vp9_sub_pixel_variance32x16_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance32x16 = vp9_sub_pixel_variance32x16_ssse3;
+
+    vp9_sub_pixel_avg_variance32x16 = vp9_sub_pixel_avg_variance32x16_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance32x16 = vp9_sub_pixel_avg_variance32x16_ssse3;
+
+    vp9_sub_pixel_variance16x32 = vp9_sub_pixel_variance16x32_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance16x32 = vp9_sub_pixel_variance16x32_ssse3;
+
+    vp9_sub_pixel_avg_variance16x32 = vp9_sub_pixel_avg_variance16x32_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance16x32 = vp9_sub_pixel_avg_variance16x32_ssse3;
+
+    vp9_sub_pixel_variance32x32 = vp9_sub_pixel_variance32x32_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance32x32 = vp9_sub_pixel_variance32x32_ssse3;
+
+    vp9_sub_pixel_avg_variance32x32 = vp9_sub_pixel_avg_variance32x32_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance32x32 = vp9_sub_pixel_avg_variance32x32_ssse3;
+
+    vp9_sub_pixel_variance16x16 = vp9_sub_pixel_variance16x16_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance16x16 = vp9_sub_pixel_variance16x16_ssse3;
+
+    vp9_sub_pixel_avg_variance16x16 = vp9_sub_pixel_avg_variance16x16_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance16x16 = vp9_sub_pixel_avg_variance16x16_ssse3;
+
+    vp9_sub_pixel_variance8x16 = vp9_sub_pixel_variance8x16_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance8x16 = vp9_sub_pixel_variance8x16_ssse3;
+
+    vp9_sub_pixel_avg_variance8x16 = vp9_sub_pixel_avg_variance8x16_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance8x16 = vp9_sub_pixel_avg_variance8x16_ssse3;
+
+    vp9_sub_pixel_variance16x8 = vp9_sub_pixel_variance16x8_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance16x8 = vp9_sub_pixel_variance16x8_ssse3;
+
+    vp9_sub_pixel_avg_variance16x8 = vp9_sub_pixel_avg_variance16x8_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance16x8 = vp9_sub_pixel_avg_variance16x8_ssse3;
+
+    vp9_sub_pixel_variance8x8 = vp9_sub_pixel_variance8x8_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance8x8 = vp9_sub_pixel_variance8x8_ssse3;
+
+    vp9_sub_pixel_avg_variance8x8 = vp9_sub_pixel_avg_variance8x8_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance8x8 = vp9_sub_pixel_avg_variance8x8_ssse3;
+
+    vp9_sub_pixel_variance8x4 = vp9_sub_pixel_variance8x4_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance8x4 = vp9_sub_pixel_variance8x4_ssse3;
+
+    vp9_sub_pixel_avg_variance8x4 = vp9_sub_pixel_avg_variance8x4_sse2;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance8x4 = vp9_sub_pixel_avg_variance8x4_ssse3;
+
+    vp9_sub_pixel_variance4x8 = vp9_sub_pixel_variance4x8_sse;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance4x8 = vp9_sub_pixel_variance4x8_ssse3;
+
+    vp9_sub_pixel_avg_variance4x8 = vp9_sub_pixel_avg_variance4x8_sse;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance4x8 = vp9_sub_pixel_avg_variance4x8_ssse3;
+
+    vp9_sub_pixel_variance4x4 = vp9_sub_pixel_variance4x4_sse;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_variance4x4 = vp9_sub_pixel_variance4x4_ssse3;
+
+    vp9_sub_pixel_avg_variance4x4 = vp9_sub_pixel_avg_variance4x4_sse;
+    if (flags & HAS_SSSE3) vp9_sub_pixel_avg_variance4x4 = vp9_sub_pixel_avg_variance4x4_ssse3;
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+    vp9_sad16x16x3 = vp9_sad16x16x3_c;
+    if (flags & HAS_SSE3) vp9_sad16x16x3 = vp9_sad16x16x3_sse3;
+    if (flags & HAS_SSSE3) vp9_sad16x16x3 = vp9_sad16x16x3_ssse3;
+
+    vp9_sad16x8x3 = vp9_sad16x8x3_c;
+    if (flags & HAS_SSE3) vp9_sad16x8x3 = vp9_sad16x8x3_sse3;
+    if (flags & HAS_SSSE3) vp9_sad16x8x3 = vp9_sad16x8x3_ssse3;
+
+    vp9_sad8x16x3 = vp9_sad8x16x3_c;
+    if (flags & HAS_SSE3) vp9_sad8x16x3 = vp9_sad8x16x3_sse3;
+
+    vp9_sad8x8x3 = vp9_sad8x8x3_c;
+    if (flags & HAS_SSE3) vp9_sad8x8x3 = vp9_sad8x8x3_sse3;
+
+    vp9_sad4x4x3 = vp9_sad4x4x3_c;
+    if (flags & HAS_SSE3) vp9_sad4x4x3 = vp9_sad4x4x3_sse3;
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+    vp9_quantize_b = vp9_quantize_b_c;
+    if (flags & HAS_SSSE3) vp9_quantize_b = vp9_quantize_b_ssse3;
+
+    vp9_quantize_b_32x32 = vp9_quantize_b_32x32_c;
+    if (flags & HAS_SSSE3) vp9_quantize_b_32x32 = vp9_quantize_b_32x32_ssse3;
+
+
+
+
+
+
+
+
+
+
+    vp9_full_search_sad = vp9_full_search_sad_c;
+    if (flags & HAS_SSE3) vp9_full_search_sad = vp9_full_search_sadx3;
+    if (flags & HAS_SSE4_1) vp9_full_search_sad = vp9_full_search_sadx8;
+
+    vp9_refining_search_sad = vp9_refining_search_sad_c;
+    if (flags & HAS_SSE3) vp9_refining_search_sad = vp9_refining_search_sadx4;
+
+    vp9_diamond_search_sad = vp9_diamond_search_sad_c;
+    if (flags & HAS_SSE3) vp9_diamond_search_sad = vp9_diamond_search_sadx4;
+}
+#endif
+#endif
diff --git a/media/libvpx/vpx_config.h b/media/libvpx/vpx_config.h
index bd370b55a1a..12116978edf 100644
--- a/media/libvpx/vpx_config.h
+++ b/media/libvpx/vpx_config.h
@@ -2,11 +2,19 @@
 
 #if defined(_WIN64)
 /* 64 bit Windows */
+#ifdef _MSC_VER
 #include "vpx_config_x86_64-win64-vs8.h"
+#else
+#include "vpx_config_x86_64-win64-gcc.h"
+#endif
 
 #elif defined(_WIN32)
 /* 32 bit Windows, MSVC. */
+#ifdef _MSC_VER
 #include "vpx_config_x86-win32-vs8.h"
+#else
+#include "vpx_config_x86-win32-gcc.h"
+#endif
 
 #elif defined(__APPLE__) && defined(__x86_64__)
 /* 64 bit MacOS. */
diff --git a/media/libvpx/vpx_config_x86-win32-gcc.asm b/media/libvpx/vpx_config_x86-win32-gcc.asm
new file mode 100644
index 00000000000..70072f689f3
--- /dev/null
+++ b/media/libvpx/vpx_config_x86-win32-gcc.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 1
+HAVE_SSSE3 equ 1
+HAVE_SSE4_1 equ 1
+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 0
+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 0
+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_MD5 equ 1
+CONFIG_DEQUANT_TOKENS equ 0
+CONFIG_DC_RECON equ 0
+CONFIG_RUNTIME_CPU_DETECT equ 1
+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 1
+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 1
+CONFIG_MULTI_RES_ENCODING equ 1
+CONFIG_TEMPORAL_DENOISING equ 1
+CONFIG_EXPERIMENTAL equ 0
+CONFIG_DECRYPT equ 0
+CONFIG_MULTIPLE_ARF equ 0
+CONFIG_NON420 equ 0
+CONFIG_ALPHA equ 0
diff --git a/media/libvpx/vpx_config_x86-win32-gcc.h b/media/libvpx/vpx_config_x86-win32-gcc.h
new file mode 100644
index 00000000000..2e271e0b465
--- /dev/null
+++ b/media/libvpx/vpx_config_x86-win32-gcc.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 1
+#define HAVE_SSSE3 1
+#define HAVE_SSE4_1 1
+#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 0
+#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 0
+#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_MD5 1
+#define CONFIG_DEQUANT_TOKENS 0
+#define CONFIG_DC_RECON 0
+#define CONFIG_RUNTIME_CPU_DETECT 1
+#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 1
+#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 1
+#define CONFIG_MULTI_RES_ENCODING 1
+#define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_DECRYPT 0
+#define CONFIG_MULTIPLE_ARF 0
+#define CONFIG_NON420 0
+#define CONFIG_ALPHA 0
+#endif /* VPX_CONFIG_H */
diff --git a/media/libvpx/vpx_config_x86_64-win64-gcc.asm b/media/libvpx/vpx_config_x86_64-win64-gcc.asm
new file mode 100644
index 00000000000..2d88f79d681
--- /dev/null
+++ b/media/libvpx/vpx_config_x86_64-win64-gcc.asm
@@ -0,0 +1,80 @@
+ARCH_ARM equ 0
+ARCH_MIPS equ 0
+ARCH_X86 equ 0
+ARCH_X86_64 equ 1
+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 1
+HAVE_SSSE3 equ 1
+HAVE_SSE4_1 equ 1
+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 0
+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 0
+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_MD5 equ 1
+CONFIG_DEQUANT_TOKENS equ 0
+CONFIG_DC_RECON equ 0
+CONFIG_RUNTIME_CPU_DETECT equ 1
+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 1
+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 1
+CONFIG_MULTI_RES_ENCODING equ 1
+CONFIG_TEMPORAL_DENOISING equ 1
+CONFIG_EXPERIMENTAL equ 0
+CONFIG_DECRYPT equ 0
+CONFIG_MULTIPLE_ARF equ 0
+CONFIG_NON420 equ 0
+CONFIG_ALPHA equ 0
diff --git a/media/libvpx/vpx_config_x86_64-win64-gcc.h b/media/libvpx/vpx_config_x86_64-win64-gcc.h
new file mode 100644
index 00000000000..8495092b226
--- /dev/null
+++ b/media/libvpx/vpx_config_x86_64-win64-gcc.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 0
+#define ARCH_X86_64 1
+#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 1
+#define HAVE_SSSE3 1
+#define HAVE_SSE4_1 1
+#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 0
+#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 0
+#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_MD5 1
+#define CONFIG_DEQUANT_TOKENS 0
+#define CONFIG_DC_RECON 0
+#define CONFIG_RUNTIME_CPU_DETECT 1
+#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 1
+#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 1
+#define CONFIG_MULTI_RES_ENCODING 1
+#define CONFIG_TEMPORAL_DENOISING 1
+#define CONFIG_EXPERIMENTAL 0
+#define CONFIG_DECRYPT 0
+#define CONFIG_MULTIPLE_ARF 0
+#define CONFIG_NON420 0
+#define CONFIG_ALPHA 0
+#endif /* VPX_CONFIG_H */
diff --git a/media/libvpx/vpx_scale_rtcd.h b/media/libvpx/vpx_scale_rtcd.h
index 6594b489edc..fd9523425b9 100644
--- a/media/libvpx/vpx_scale_rtcd.h
+++ b/media/libvpx/vpx_scale_rtcd.h
@@ -8,11 +8,19 @@
 
 #if defined(_WIN64)
 /* 64 bit Windows */
+#ifdef _MSC_VER
 #include "vpx_scale_rtcd_x86_64-win64-vs8.h"
+#else
+#include "vpx_scale_rtcd_x86_64-win64-gcc.h"
+#endif
 
 #elif defined(_WIN32)
 /* 32 bit Windows, MSVC. */
+#ifdef _MSC_VER
 #include "vpx_scale_rtcd_x86-win32-vs8.h"
+#else
+#include "vpx_scale_rtcd_x86-win32-gcc.h"
+#endif
 
 #elif defined(__APPLE__) && defined(__x86_64__)
 /* 64 bit MacOS. */
diff --git a/media/libvpx/vpx_scale_rtcd_x86-win32-gcc.h b/media/libvpx/vpx_scale_rtcd_x86-win32-gcc.h
new file mode 100644
index 00000000000..e7113836607
--- /dev/null
+++ b/media/libvpx/vpx_scale_rtcd_x86-win32-gcc.h
@@ -0,0 +1,61 @@
+#ifndef VPX_SCALE_RTCD_H_
+#define VPX_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+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_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_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_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_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_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_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_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 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 vp9_extend_frame_borders_c(struct yv12_buffer_config *ybf, int subsampling_x, int subsampling_y);
+#define vp9_extend_frame_borders vp9_extend_frame_borders_c
+
+void vp9_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf, int subsampling_x, int subsampling_y);
+#define vp9_extend_frame_inner_borders vp9_extend_frame_inner_borders_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
+#endif
diff --git a/media/libvpx/vpx_scale_rtcd_x86_64-win64-gcc.h b/media/libvpx/vpx_scale_rtcd_x86_64-win64-gcc.h
new file mode 100644
index 00000000000..e7113836607
--- /dev/null
+++ b/media/libvpx/vpx_scale_rtcd_x86_64-win64-gcc.h
@@ -0,0 +1,61 @@
+#ifndef VPX_SCALE_RTCD_H_
+#define VPX_SCALE_RTCD_H_
+
+#ifdef RTCD_C
+#define RTCD_EXTERN
+#else
+#define RTCD_EXTERN extern
+#endif
+
+struct yv12_buffer_config;
+
+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_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_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_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_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_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_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
+#define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_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 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 vp9_extend_frame_borders_c(struct yv12_buffer_config *ybf, int subsampling_x, int subsampling_y);
+#define vp9_extend_frame_borders vp9_extend_frame_borders_c
+
+void vp9_extend_frame_inner_borders_c(struct yv12_buffer_config *ybf, int subsampling_x, int subsampling_y);
+#define vp9_extend_frame_inner_borders vp9_extend_frame_inner_borders_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
+#endif