libvpx: Pull from upstream

source/config/mac/ia32/vp9_rtcd.h

 #ifndef VP__RTCD_
 #define VP__RTCD_
 
 #ifdef RTCD_C
 #define RTCD_EXTERN
 #else
 #define RTCD_EXTERN extern
 #endif
 
 /*
  * VP9
  */
 
+#include "vpx/vpx_integer.h"
+
 struct loop_filter_info;
 struct blockd;
 struct macroblockd;
 struct loop_filter_info;
 
 /* Encoder forward decls */
 struct block;
 struct macroblock;
 struct variance_vtable;
 
 #define DEC_MVCOSTS int *mvjcost, int *mvcost[2]
 union int_mv;
 struct yv12_buffer_config;
 
-void vp9_filter_block2d_4x4_8_c(const unsigned char *src_ptr, const unsigned int src_stride, const short *HFilter_aligned16, const short *VFilter_aligned16, unsigned char *dst_ptr, unsigned int dst_stride);
-void vp9_filter_block2d_4x4_8_sse2(const unsigned char *src_ptr, const unsigned int src_stride, const short *HFilter_aligned16, const short *VFilter_aligned16, unsigned char *dst_ptr, unsigned int dst_stride);
-void vp9_filter_block2d_4x4_8_sse4_1(const unsigned char *src_ptr, const unsigned int src_stride, const short *HFilter_aligned16, const short *VFilter_aligned16, unsigned char *dst_ptr, unsigned int dst_stride);
-RTCD_EXTERN void (*vp9_filter_block2d_4x4_8)(const unsigned char *src_ptr, const unsigned int src_stride, const short *HFilter_aligned16, const short *VFilter_aligned16, unsigned char *dst_ptr, unsigned int dst_stride);
+void vp9_filter_block2d_4x4_8_c(const uint8_t *src_ptr, const unsigned int src_stride, const int16_t *HFilter_aligned16, const int16_t *VFilter_aligned16, uint8_t *dst_ptr, unsigned int dst_stride);
+void vp9_filter_block2d_4x4_8_sse2(const uint8_t *src_ptr, const unsigned int src_stride, const int16_t *HFilter_aligned16, const int16_t *VFilter_aligned16, uint8_t *dst_ptr, unsigned int dst_stride);
+void vp9_filter_block2d_4x4_8_sse4_1(const uint8_t *src_ptr, const unsigned int src_stride, const int16_t *HFilter_aligned16, const int16_t *VFilter_aligned16, uint8_t *dst_ptr, unsigned int dst_stride);
+RTCD_EXTERN void (*vp9_filter_block2d_4x4_8)(const uint8_t *src_ptr, const unsigned int src_stride, const int16_t *HFilter_aligned16, const int16_t *VFilter_aligned16, uint8_t *dst_ptr, unsigned int dst_stride);
 
-void vp9_filter_block2d_8x4_8_c(const unsigned char *src_ptr, const unsigned int src_stride, const short *HFilter_aligned16, const short *VFilter_aligned16, unsigned char *dst_ptr, unsigned int dst_stride);
-void vp9_filter_block2d_8x4_8_ssse3(const unsigned char *src_ptr, const unsigned int src_stride, const short *HFilter_aligned16, const short *VFilter_aligned16, unsigned char *dst_ptr, unsigned int dst_stride);
-RTCD_EXTERN void (*vp9_filter_block2d_8x4_8)(const unsigned char *src_ptr, const unsigned int src_stride, const short *HFilter_aligned16, const short *VFilter_aligned16, unsigned char *dst_ptr, unsigned int dst_stride);
+void vp9_filter_block2d_8x4_8_c(const uint8_t *src_ptr, const unsigned int src_stride, const int16_t *HFilter_aligned16, const int16_t *VFilter_aligned16, uint8_t *dst_ptr, unsigned int dst_stride);
+void vp9_filter_block2d_8x4_8_ssse3(const uint8_t *src_ptr, const unsigned int src_stride, const int16_t *HFilter_aligned16, const int16_t *VFilter_aligned16, uint8_t *dst_ptr, unsigned int dst_stride);
+RTCD_EXTERN void (*vp9_filter_block2d_8x4_8)(const uint8_t *src_ptr, const unsigned int src_stride, const int16_t *HFilter_aligned16, const int16_t *VFilter_aligned16, uint8_t *dst_ptr, unsigned int dst_stride);
 
-void vp9_filter_block2d_8x8_8_c(const unsigned char *src_ptr, const unsigned int src_stride, const short *HFilter_aligned16, const short *VFilter_aligned16, unsigned char *dst_ptr, unsigned int dst_stride);
-void vp9_filter_block2d_8x8_8_ssse3(const unsigned char *src_ptr, const unsigned int src_stride, const short *HFilter_aligned16, const short *VFilter_aligned16, unsigned char *dst_ptr, unsigned int dst_stride);
-RTCD_EXTERN void (*vp9_filter_block2d_8x8_8)(const unsigned char *src_ptr, const unsigned int src_stride, const short *HFilter_aligned16, const short *VFilter_aligned16, unsigned char *dst_ptr, unsigned int dst_stride);
+void vp9_filter_block2d_8x8_8_c(const uint8_t *src_ptr, const unsigned int src_stride, const int16_t *HFilter_aligned16, const int16_t *VFilter_aligned16, uint8_t *dst_ptr, unsigned int dst_stride);
+void vp9_filter_block2d_8x8_8_ssse3(const uint8_t *src_ptr, const unsigned int src_stride, const int16_t *HFilter_aligned16, const int16_t *VFilter_aligned16, uint8_t *dst_ptr, unsigned int dst_stride);
+RTCD_EXTERN void (*vp9_filter_block2d_8x8_8)(const uint8_t *src_ptr, const unsigned int src_stride, const int16_t *HFilter_aligned16, const int16_t *VFilter_aligned16, uint8_t *dst_ptr, unsigned int dst_stride);
 
-void vp9_filter_block2d_16x16_8_c(const unsigned char *src_ptr, const unsigned int src_stride, const short *HFilter_aligned16, const short *VFilter_aligned16, unsigned char *dst_ptr, unsigned int dst_stride);
-void vp9_filter_block2d_16x16_8_ssse3(const unsigned char *src_ptr, const unsigned int src_stride, const short *HFilter_aligned16, const short *VFilter_aligned16, unsigned char *dst_ptr, unsigned int dst_stride);
-RTCD_EXTERN void (*vp9_filter_block2d_16x16_8)(const unsigned char *src_ptr, const unsigned int src_stride, const short *HFilter_aligned16, const short *VFilter_aligned16, unsigned char *dst_ptr, unsigned int dst_stride);
+void vp9_filter_block2d_16x16_8_c(const uint8_t *src_ptr, const unsigned int src_stride, const int16_t *HFilter_aligned16, const int16_t *VFilter_aligned16, uint8_t *dst_ptr, unsigned int dst_stride);
+void vp9_filter_block2d_16x16_8_ssse3(const uint8_t *src_ptr, const unsigned int src_stride, const int16_t *HFilter_aligned16, const int16_t *VFilter_aligned16, uint8_t *dst_ptr, unsigned int dst_stride);
+RTCD_EXTERN void (*vp9_filter_block2d_16x16_8)(const uint8_t *src_ptr, const unsigned int src_stride, const int16_t *HFilter_aligned16, const int16_t *VFilter_aligned16, uint8_t *dst_ptr, unsigned int dst_stride);
 
 void vp9_dequantize_b_c(struct blockd *x);
 #define vp9_dequantize_b vp9_dequantize_b_c
 
 void vp9_dequantize_b_2x2_c(struct blockd *x);
 #define vp9_dequantize_b_2x2 vp9_dequantize_b_2x2_c
 
-void vp9_dequant_dc_idct_add_y_block_8x8_c(short *q, const short *dq, unsigned char *pre, unsigned char *dst, int stride, unsigned short *eobs, const short *dc, struct macroblockd *xd);
+void vp9_dequant_dc_idct_add_y_block_8x8_c(int16_t *q, const int16_t *dq, uint8_t *pre, uint8_t *dst, int stride, uint16_t *eobs, const int16_t *dc, struct macroblockd *xd);
 #define vp9_dequant_dc_idct_add_y_block_8x8 vp9_dequant_dc_idct_add_y_block_8x8_c
 
-void vp9_dequant_idct_add_y_block_8x8_c(short *q, const short *dq, unsigned char *pre, unsigned char *dst, int stride, unsigned short *eobs, struct macroblockd *xd);
+void vp9_dequant_idct_add_y_block_8x8_c(int16_t *q, const int16_t *dq, uint8_t *pre, uint8_t *dst, int stride, uint16_t *eobs, struct macroblockd *xd);
 #define vp9_dequant_idct_add_y_block_8x8 vp9_dequant_idct_add_y_block_8x8_c
 
-void vp9_dequant_idct_add_uv_block_8x8_c(short *q, const short *dq, unsigned char *pre, unsigned char *dstu, unsigned char *dstv, int stride, unsigned short *eobs, struct macroblockd *xd);
+void vp9_dequant_idct_add_uv_block_8x8_c(int16_t *q, const int16_t *dq, uint8_t *pre, uint8_t *dstu, uint8_t *dstv, int stride, uint16_t *eobs, struct macroblockd *xd);
 #define vp9_dequant_idct_add_uv_block_8x8 vp9_dequant_idct_add_uv_block_8x8_c
 
-void vp9_dequant_idct_add_16x16_c(short *input, const short *dq, unsigned char *pred, unsigned char *dest, int pitch, int stride, unsigned short eobs);
+void vp9_dequant_idct_add_16x16_c(int16_t *input, const int16_t *dq, uint8_t *pred, uint8_t *dest, int pitch, int stride, int eob);
 #define vp9_dequant_idct_add_16x16 vp9_dequant_idct_add_16x16_c
 
-void vp9_dequant_idct_add_8x8_c(short *input, const short *dq, unsigned char *pred, unsigned char *dest, int pitch, int stride, int dc, unsigned short eobs);
+void vp9_dequant_idct_add_8x8_c(int16_t *input, const int16_t *dq, uint8_t *pred, uint8_t *dest, int pitch, int stride, int dc, int eob);
 #define vp9_dequant_idct_add_8x8 vp9_dequant_idct_add_8x8_c
 
-void vp9_dequant_idct_add_c(short *input, const short *dq, unsigned char *pred, unsigned char *dest, int pitch, int stride);
+void vp9_dequant_idct_add_c(int16_t *input, const int16_t *dq, uint8_t *pred, uint8_t *dest, int pitch, int stride);
 #define vp9_dequant_idct_add vp9_dequant_idct_add_c
 
-void vp9_dequant_dc_idct_add_c(short *input, const short *dq, unsigned char *pred, unsigned char *dest, int pitch, int stride, int Dc);
+void vp9_dequant_dc_idct_add_c(int16_t *input, const int16_t *dq, uint8_t *pred, uint8_t *dest, int pitch, int stride, int dc);
 #define vp9_dequant_dc_idct_add vp9_dequant_dc_idct_add_c
 
-void vp9_dequant_dc_idct_add_y_block_c(short *q, const short *dq, unsigned char *pre, unsigned char *dst, int stride, unsigned short *eobs, const short *dc);
+void vp9_dequant_dc_idct_add_y_block_c(int16_t *q, const int16_t *dq, uint8_t *pre, uint8_t *dst, int stride, uint16_t *eobs, const int16_t *dcs);
 #define vp9_dequant_dc_idct_add_y_block vp9_dequant_dc_idct_add_y_block_c
 
-void vp9_dequant_idct_add_y_block_c(short *q, const short *dq, unsigned char *pre, unsigned char *dst, int stride, unsigned short *eobs);
+void vp9_dequant_idct_add_y_block_c(int16_t *q, const int16_t *dq, uint8_t *pre, uint8_t *dst, int stride, uint16_t *eobs);
 #define vp9_dequant_idct_add_y_block vp9_dequant_idct_add_y_block_c
 
-void vp9_dequant_idct_add_uv_block_c(short *q, const short *dq, unsigned char *pre, unsigned char *dstu, unsigned char *dstv, int stride, unsigned short *eobs);
+void vp9_dequant_idct_add_uv_block_c(int16_t *q, const int16_t *dq, uint8_t *pre, uint8_t *dstu, uint8_t *dstv, int stride, uint16_t *eobs);
 #define vp9_dequant_idct_add_uv_block vp9_dequant_idct_add_uv_block_c
 
-void vp9_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
-void vp9_copy_mem16x16_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
-void vp9_copy_mem16x16_sse2(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
-RTCD_EXTERN void (*vp9_copy_mem16x16)(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp9_dequant_idct_add_32x32_c(int16_t *q, const int16_t *dq, uint8_t *pre, uint8_t *dst, int pitch, int stride, int eob);
+#define vp9_dequant_idct_add_32x32 vp9_dequant_idct_add_32x32_c
 
-void vp9_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
-void vp9_copy_mem8x8_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
-RTCD_EXTERN void (*vp9_copy_mem8x8)(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp9_dequant_idct_add_uv_block_16x16_c(int16_t *q, const int16_t *dq, uint8_t *dstu, uint8_t *dstv, int stride, uint16_t *eobs);
+#define vp9_dequant_idct_add_uv_block_16x16 vp9_dequant_idct_add_uv_block_16x16_c
 
-void vp9_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
-void vp9_copy_mem8x4_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
-RTCD_EXTERN void (*vp9_copy_mem8x4)(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp9_copy_mem16x16_c(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
+void vp9_copy_mem16x16_mmx(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
+void vp9_copy_mem16x16_sse2(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
+RTCD_EXTERN void (*vp9_copy_mem16x16)(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
 
-void vp9_avg_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp9_copy_mem8x8_c(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
+void vp9_copy_mem8x8_mmx(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
+RTCD_EXTERN void (*vp9_copy_mem8x8)(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
+
+void vp9_copy_mem8x4_c(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
+void vp9_copy_mem8x4_mmx(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
+RTCD_EXTERN void (*vp9_copy_mem8x4)(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
+
+void vp9_avg_mem16x16_c(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
 #define vp9_avg_mem16x16 vp9_avg_mem16x16_c
 
-void vp9_avg_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp9_avg_mem8x8_c(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
 #define vp9_avg_mem8x8 vp9_avg_mem8x8_c
 
-void vp9_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
-void vp9_copy_mem8x4_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
-RTCD_EXTERN void (*vp9_copy_mem8x4)(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
+void vp9_copy_mem8x4_c(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
+void vp9_copy_mem8x4_mmx(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
+RTCD_EXTERN void (*vp9_copy_mem8x4)(uint8_t *src, int src_pitch, uint8_t *dst, int dst_pitch);
 
-void vp9_recon_b_c(unsigned char *pred_ptr, short *diff_ptr, unsigned char *dst_ptr, int stride);
+void vp9_recon_b_c(uint8_t *pred_ptr, int16_t *diff_ptr, uint8_t *dst_ptr, int stride);
 #define vp9_recon_b vp9_recon_b_c
 
-void vp9_recon_uv_b_c(unsigned char *pred_ptr, short *diff_ptr, unsigned char *dst_ptr, int stride);
+void vp9_recon_uv_b_c(uint8_t *pred_ptr, int16_t *diff_ptr, uint8_t *dst_ptr, int stride);
 #define vp9_recon_uv_b vp9_recon_uv_b_c
 
-void vp9_recon2b_c(unsigned char *pred_ptr, short *diff_ptr, unsigned char *dst_ptr, int stride);
-void vp9_recon2b_sse2(unsigned char *pred_ptr, short *diff_ptr, unsigned char *dst_ptr, int stride);
-RTCD_EXTERN void (*vp9_recon2b)(unsigned char *pred_ptr, short *diff_ptr, unsigned char *dst_ptr, int stride);
+void vp9_recon2b_c(uint8_t *pred_ptr, int16_t *diff_ptr, uint8_t *dst_ptr, int stride);
+void vp9_recon2b_sse2(uint8_t *pred_ptr, int16_t *diff_ptr, uint8_t *dst_ptr, int stride);
+RTCD_EXTERN void (*vp9_recon2b)(uint8_t *pred_ptr, int16_t *diff_ptr, uint8_t *dst_ptr, int stride);
 
-void vp9_recon4b_c(unsigned char *pred_ptr, short *diff_ptr, unsigned char *dst_ptr, int stride);
-void vp9_recon4b_sse2(unsigned char *pred_ptr, short *diff_ptr, unsigned char *dst_ptr, int stride);
-RTCD_EXTERN void (*vp9_recon4b)(unsigned char *pred_ptr, short *diff_ptr, unsigned char *dst_ptr, int stride);
+void vp9_recon4b_c(uint8_t *pred_ptr, int16_t *diff_ptr, uint8_t *dst_ptr, int stride);
+void vp9_recon4b_sse2(uint8_t *pred_ptr, int16_t *diff_ptr, uint8_t *dst_ptr, int stride);
+RTCD_EXTERN void (*vp9_recon4b)(uint8_t *pred_ptr, int16_t *diff_ptr, uint8_t *dst_ptr, int stride);
 
 void vp9_recon_mb_c(struct macroblockd *x);
 #define vp9_recon_mb vp9_recon_mb_c
 
 void vp9_recon_mby_c(struct macroblockd *x);
 #define vp9_recon_mby vp9_recon_mby_c
 
-void vp9_recon_mby_s_c(struct macroblockd *x, unsigned char *dst);
+void vp9_recon_mby_s_c(struct macroblockd *x, uint8_t *dst);
 #define vp9_recon_mby_s vp9_recon_mby_s_c
 
-void vp9_recon_mbuv_s_c(struct macroblockd *x, unsigned char *udst, unsigned char *vdst);
+void vp9_recon_mbuv_s_c(struct macroblockd *x, uint8_t *udst, uint8_t *vdst);
 #define vp9_recon_mbuv_s vp9_recon_mbuv_s_c
 
+void vp9_recon_sby_s_c(struct macroblockd *x, uint8_t *dst);
+#define vp9_recon_sby_s vp9_recon_sby_s_c
+
+void vp9_recon_sbuv_s_c(struct macroblockd *x, uint8_t *udst, uint8_t *vdst);
+#define vp9_recon_sbuv_s vp9_recon_sbuv_s_c
+
 void vp9_build_intra_predictors_mby_s_c(struct macroblockd *x);
 #define vp9_build_intra_predictors_mby_s vp9_build_intra_predictors_mby_s_c
 
 void vp9_build_intra_predictors_sby_s_c(struct macroblockd *x);
 #define vp9_build_intra_predictors_sby_s vp9_build_intra_predictors_sby_s_c
 
 void vp9_build_intra_predictors_sbuv_s_c(struct macroblockd *x);
 #define vp9_build_intra_predictors_sbuv_s vp9_build_intra_predictors_sbuv_s_c
 
 void vp9_build_intra_predictors_mby_c(struct macroblockd *x);
 #define vp9_build_intra_predictors_mby vp9_build_intra_predictors_mby_c
 
-void vp9_build_comp_intra_predictors_mby_c(struct macroblockd *x);
-#define vp9_build_comp_intra_predictors_mby vp9_build_comp_intra_predictors_mby_c
-
 void vp9_build_intra_predictors_mby_s_c(struct macroblockd *x);
 #define vp9_build_intra_predictors_mby_s vp9_build_intra_predictors_mby_s_c
 
 void vp9_build_intra_predictors_mbuv_c(struct macroblockd *x);
 #define vp9_build_intra_predictors_mbuv vp9_build_intra_predictors_mbuv_c
 
 void vp9_build_intra_predictors_mbuv_s_c(struct macroblockd *x);
 #define vp9_build_intra_predictors_mbuv_s vp9_build_intra_predictors_mbuv_s_c
 
-void vp9_build_comp_intra_predictors_mbuv_c(struct macroblockd *x);
-#define vp9_build_comp_intra_predictors_mbuv vp9_build_comp_intra_predictors_mbuv_c
-
-void vp9_intra4x4_predict_c(struct blockd *x, int b_mode, unsigned char *predictor);
+void vp9_build_intra_predictors_sb64y_s_c(struct macroblockd *x);
+#define vp9_build_intra_predictors_sb64y_s vp9_build_intra_predictors_sb64y_s_c
+
+void vp9_build_intra_predictors_sb64uv_s_c(struct macroblockd *x);
+#define vp9_build_intra_predictors_sb64uv_s vp9_build_intra_predictors_sb64uv_s_c
+
+void vp9_intra4x4_predict_c(struct blockd *x, int b_mode, uint8_t *predictor);
 #define vp9_intra4x4_predict vp9_intra4x4_predict_c
 
-void vp9_comp_intra4x4_predict_c(struct blockd *x, int b_mode, int second_mode, unsigned char *predictor);
-#define vp9_comp_intra4x4_predict vp9_comp_intra4x4_predict_c
-
-void vp9_intra8x8_predict_c(struct blockd *x, int b_mode, unsigned char *predictor);
+void vp9_intra8x8_predict_c(struct blockd *x, int b_mode, uint8_t *predictor);
 #define vp9_intra8x8_predict vp9_intra8x8_predict_c
 
-void vp9_comp_intra8x8_predict_c(struct blockd *x, int b_mode, int second_mode, unsigned char *predictor);
-#define vp9_comp_intra8x8_predict vp9_comp_intra8x8_predict_c
-
-void vp9_intra_uv4x4_predict_c(struct blockd *x, int b_mode, unsigned char *predictor);
+void vp9_intra_uv4x4_predict_c(struct blockd *x, int b_mode, uint8_t *predictor);
 #define vp9_intra_uv4x4_predict vp9_intra_uv4x4_predict_c
 
-void vp9_comp_intra_uv4x4_predict_c(struct blockd *x, int b_mode, int second_mode, unsigned char *predictor);
-#define vp9_comp_intra_uv4x4_predict vp9_comp_intra_uv4x4_predict_c
-
-void vp9_loop_filter_mbv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
-void vp9_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 (*vp9_loop_filter_mbv)(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
-
-void vp9_loop_filter_bv_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
-void vp9_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 (*vp9_loop_filter_bv)(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
-
-void vp9_loop_filter_bv8x8_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
-void vp9_loop_filter_bv8x8_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
-RTCD_EXTERN void (*vp9_loop_filter_bv8x8)(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
-
-void vp9_loop_filter_mbh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
-void vp9_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 (*vp9_loop_filter_mbh)(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
-
-void vp9_loop_filter_bh_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
-void vp9_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 (*vp9_loop_filter_bh)(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
-
-void vp9_loop_filter_bh8x8_c(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
-void vp9_loop_filter_bh8x8_sse2(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
-RTCD_EXTERN void (*vp9_loop_filter_bh8x8)(unsigned char *y, unsigned char *u, unsigned char *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
-
-void vp9_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
-void vp9_loop_filter_simple_vertical_edge_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
-void vp9_loop_filter_simple_vertical_edge_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
-RTCD_EXTERN void (*vp9_loop_filter_simple_mbv)(unsigned char *y, int ystride, const unsigned char *blimit);
-
-void vp9_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
-void vp9_loop_filter_simple_horizontal_edge_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
-void vp9_loop_filter_simple_horizontal_edge_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
-RTCD_EXTERN void (*vp9_loop_filter_simple_mbh)(unsigned char *y, int ystride, const unsigned char *blimit);
-
-void vp9_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
-void vp9_loop_filter_bvs_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
-void vp9_loop_filter_bvs_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
-RTCD_EXTERN void (*vp9_loop_filter_simple_bv)(unsigned char *y, int ystride, const unsigned char *blimit);
-
-void vp9_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
-void vp9_loop_filter_bhs_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
-void vp9_loop_filter_bhs_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
-RTCD_EXTERN void (*vp9_loop_filter_simple_bh)(unsigned char *y, int ystride, const unsigned char *blimit);
-
-void vp9_mbpost_proc_down_c(unsigned char *dst, int pitch, int rows, int cols, int flimit);
-void vp9_mbpost_proc_down_mmx(unsigned char *dst, int pitch, int rows, int cols, int flimit);
-void vp9_mbpost_proc_down_xmm(unsigned char *dst, int pitch, int rows, int cols, int flimit);
-RTCD_EXTERN void (*vp9_mbpost_proc_down)(unsigned char *dst, int pitch, int rows, int cols, int flimit);
-
-void vp9_mbpost_proc_across_ip_c(unsigned char *src, int pitch, int rows, int cols, int flimit);
-void vp9_mbpost_proc_across_ip_xmm(unsigned char *src, int pitch, int rows, int cols, int flimit);
-RTCD_EXTERN void (*vp9_mbpost_proc_across_ip)(unsigned char *src, int pitch, int rows, int cols, int flimit);
-
-void vp9_post_proc_down_and_across_c(unsigned char *src_ptr, unsigned char *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit);
-void vp9_post_proc_down_and_across_mmx(unsigned char *src_ptr, unsigned char *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit);
-void vp9_post_proc_down_and_across_xmm(unsigned char *src_ptr, unsigned char *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit);
-RTCD_EXTERN void (*vp9_post_proc_down_and_across)(unsigned char *src_ptr, unsigned char *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit);
-
-void vp9_plane_add_noise_c(unsigned char *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
-void vp9_plane_add_noise_mmx(unsigned char *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
-void vp9_plane_add_noise_wmt(unsigned char *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
-RTCD_EXTERN void (*vp9_plane_add_noise)(unsigned char *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
-
-void vp9_blend_mb_inner_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
+void vp9_loop_filter_mbv_c(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp9_loop_filter_mbv_sse2(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+RTCD_EXTERN void (*vp9_loop_filter_mbv)(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+
+void vp9_loop_filter_bv_c(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp9_loop_filter_bv_sse2(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+RTCD_EXTERN void (*vp9_loop_filter_bv)(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+
+void vp9_loop_filter_bv8x8_c(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp9_loop_filter_bv8x8_sse2(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+RTCD_EXTERN void (*vp9_loop_filter_bv8x8)(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+
+void vp9_loop_filter_mbh_c(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp9_loop_filter_mbh_sse2(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+RTCD_EXTERN void (*vp9_loop_filter_mbh)(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+
+void vp9_loop_filter_bh_c(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp9_loop_filter_bh_sse2(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+RTCD_EXTERN void (*vp9_loop_filter_bh)(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+
+void vp9_loop_filter_bh8x8_c(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+void vp9_loop_filter_bh8x8_sse2(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+RTCD_EXTERN void (*vp9_loop_filter_bh8x8)(uint8_t *y, uint8_t *u, uint8_t *v, int ystride, int uv_stride, struct loop_filter_info *lfi);
+
+void vp9_loop_filter_simple_vertical_edge_c(uint8_t *y, int ystride, const uint8_t *blimit);
+void vp9_loop_filter_simple_vertical_edge_mmx(uint8_t *y, int ystride, const uint8_t *blimit);
+void vp9_loop_filter_simple_vertical_edge_sse2(uint8_t *y, int ystride, const uint8_t *blimit);
+RTCD_EXTERN void (*vp9_loop_filter_simple_mbv)(uint8_t *y, int ystride, const uint8_t *blimit);
+
+void vp9_loop_filter_simple_horizontal_edge_c(uint8_t *y, int ystride, const uint8_t *blimit);
+void vp9_loop_filter_simple_horizontal_edge_mmx(uint8_t *y, int ystride, const uint8_t *blimit);
+void vp9_loop_filter_simple_horizontal_edge_sse2(uint8_t *y, int ystride, const uint8_t *blimit);
+RTCD_EXTERN void (*vp9_loop_filter_simple_mbh)(uint8_t *y, int ystride, const uint8_t *blimit);
+
+void vp9_loop_filter_bvs_c(uint8_t *y, int ystride, const uint8_t *blimit);
+void vp9_loop_filter_bvs_mmx(uint8_t *y, int ystride, const uint8_t *blimit);
+void vp9_loop_filter_bvs_sse2(uint8_t *y, int ystride, const uint8_t *blimit);
+RTCD_EXTERN void (*vp9_loop_filter_simple_bv)(uint8_t *y, int ystride, const uint8_t *blimit);
+
+void vp9_loop_filter_bhs_c(uint8_t *y, int ystride, const uint8_t *blimit);
+void vp9_loop_filter_bhs_mmx(uint8_t *y, int ystride, const uint8_t *blimit);
+void vp9_loop_filter_bhs_sse2(uint8_t *y, int ystride, const uint8_t *blimit);
+RTCD_EXTERN void (*vp9_loop_filter_simple_bh)(uint8_t *y, int ystride, const uint8_t *blimit);
+
+void vp9_lpf_mbh_w_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi);
+void vp9_lpf_mbh_w_sse2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi);
+RTCD_EXTERN void (*vp9_lpf_mbh_w)(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi);
+
+void vp9_lpf_mbv_w_c(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi);
+void vp9_lpf_mbv_w_sse2(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi);
+RTCD_EXTERN void (*vp9_lpf_mbv_w)(unsigned char *y_ptr, unsigned char *u_ptr, unsigned char *v_ptr, int y_stride, int uv_stride, struct loop_filter_info *lfi);
+
+void vp9_mbpost_proc_down_c(uint8_t *dst, int pitch, int rows, int cols, int flimit);
+void vp9_mbpost_proc_down_mmx(uint8_t *dst, int pitch, int rows, int cols, int flimit);
+void vp9_mbpost_proc_down_xmm(uint8_t *dst, int pitch, int rows, int cols, int flimit);
+RTCD_EXTERN void (*vp9_mbpost_proc_down)(uint8_t *dst, int pitch, int rows, int cols, int flimit);
+
+void vp9_mbpost_proc_across_ip_c(uint8_t *src, int pitch, int rows, int cols, int flimit);
+void vp9_mbpost_proc_across_ip_xmm(uint8_t *src, int pitch, int rows, int cols, int flimit);
+RTCD_EXTERN void (*vp9_mbpost_proc_across_ip)(uint8_t *src, int pitch, int rows, int cols, int flimit);
+
+void vp9_post_proc_down_and_across_c(uint8_t *src_ptr, uint8_t *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit);
+void vp9_post_proc_down_and_across_mmx(uint8_t *src_ptr, uint8_t *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit);
+void vp9_post_proc_down_and_across_xmm(uint8_t *src_ptr, uint8_t *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit);
+RTCD_EXTERN void (*vp9_post_proc_down_and_across)(uint8_t *src_ptr, uint8_t *dst_ptr, int src_pixels_per_line, int dst_pixels_per_line, int rows, int cols, int flimit);
+
+void vp9_plane_add_noise_c(uint8_t *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
+void vp9_plane_add_noise_mmx(uint8_t *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
+void vp9_plane_add_noise_wmt(uint8_t *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
+RTCD_EXTERN void (*vp9_plane_add_noise)(uint8_t *Start, char *noise, char blackclamp[16], char whiteclamp[16], char bothclamp[16], unsigned int Width, unsigned int Height, int Pitch);
+
+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(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
+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(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
+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
 
-unsigned int vp9_sad16x3_c(const unsigned char *src_ptr, int  src_stride, const unsigned char *ref_ptr, int ref_stride);
-unsigned int vp9_sad16x3_sse2(const unsigned char *src_ptr, int  src_stride, const unsigned char *ref_ptr, int ref_stride);
-RTCD_EXTERN unsigned int (*vp9_sad16x3)(const unsigned char *src_ptr, int  src_stride, const unsigned char *ref_ptr, int ref_stride);
-
-unsigned int vp9_sad3x16_c(const unsigned char *src_ptr, int  src_stride, const unsigned char *ref_ptr, int ref_stride);
-unsigned int vp9_sad3x16_sse2(const unsigned char *src_ptr, int  src_stride, const unsigned char *ref_ptr, int ref_stride);
-RTCD_EXTERN unsigned int (*vp9_sad3x16)(const unsigned char *src_ptr, int  src_stride, const unsigned char *ref_ptr, int ref_stride);
-
-void vp9_eighttap_predict16x16_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+unsigned int vp9_sad16x3_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vp9_sad16x3_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*vp9_sad16x3)(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int vp9_sad3x16_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int ref_stride);
+unsigned int vp9_sad3x16_sse2(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int ref_stride);
+RTCD_EXTERN unsigned int (*vp9_sad3x16)(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int ref_stride);
+
+unsigned int vp9_sub_pixel_variance16x2_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int Refstride, unsigned int *sse);
+unsigned int vp9_sub_pixel_variance16x2_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int Refstride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*vp9_sub_pixel_variance16x2)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int Refstride, unsigned int *sse);
+
+void vp9_eighttap_predict16x16_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_eighttap_predict16x16 vp9_eighttap_predict16x16_c
 
-void vp9_eighttap_predict8x8_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_eighttap_predict8x8_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_eighttap_predict8x8 vp9_eighttap_predict8x8_c
 
-void vp9_eighttap_predict_avg16x16_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_eighttap_predict_avg16x16_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_eighttap_predict_avg16x16 vp9_eighttap_predict_avg16x16_c
 
-void vp9_eighttap_predict_avg8x8_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_eighttap_predict_avg8x8_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_eighttap_predict_avg8x8 vp9_eighttap_predict_avg8x8_c
 
-void vp9_eighttap_predict_avg4x4_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_eighttap_predict_avg4x4_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_eighttap_predict_avg4x4 vp9_eighttap_predict_avg4x4_c
 
-void vp9_eighttap_predict8x4_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_eighttap_predict8x4_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_eighttap_predict8x4 vp9_eighttap_predict8x4_c
 
-void vp9_eighttap_predict_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-#define vp9_eighttap_predict vp9_eighttap_predict_c
-
-void vp9_eighttap_predict16x16_sharp_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_eighttap_predict4x4_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+#define vp9_eighttap_predict4x4 vp9_eighttap_predict4x4_c
+
+void vp9_eighttap_predict16x16_sharp_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_eighttap_predict16x16_sharp vp9_eighttap_predict16x16_sharp_c
 
-void vp9_eighttap_predict8x8_sharp_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_eighttap_predict8x8_sharp_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_eighttap_predict8x8_sharp vp9_eighttap_predict8x8_sharp_c
 
-void vp9_eighttap_predict_avg16x16_sharp_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_eighttap_predict_avg16x16_sharp_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_eighttap_predict_avg16x16_sharp vp9_eighttap_predict_avg16x16_sharp_c
 
-void vp9_eighttap_predict_avg8x8_sharp_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_eighttap_predict_avg8x8_sharp_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_eighttap_predict_avg8x8_sharp vp9_eighttap_predict_avg8x8_sharp_c
 
-void vp9_eighttap_predict_avg4x4_sharp_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_eighttap_predict_avg4x4_sharp_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_eighttap_predict_avg4x4_sharp vp9_eighttap_predict_avg4x4_sharp_c
 
-void vp9_eighttap_predict8x4_sharp_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_eighttap_predict8x4_sharp_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_eighttap_predict8x4_sharp vp9_eighttap_predict8x4_sharp_c
 
-void vp9_eighttap_predict_sharp_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-#define vp9_eighttap_predict_sharp vp9_eighttap_predict_sharp_c
-
-void vp9_sixtap_predict16x16_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_eighttap_predict4x4_sharp_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+#define vp9_eighttap_predict4x4_sharp vp9_eighttap_predict4x4_sharp_c
+
+void vp9_eighttap_predict16x16_smooth_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+#define vp9_eighttap_predict16x16_smooth vp9_eighttap_predict16x16_smooth_c
+
+void vp9_eighttap_predict8x8_smooth_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+#define vp9_eighttap_predict8x8_smooth vp9_eighttap_predict8x8_smooth_c
+
+void vp9_eighttap_predict_avg16x16_smooth_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+#define vp9_eighttap_predict_avg16x16_smooth vp9_eighttap_predict_avg16x16_smooth_c
+
+void vp9_eighttap_predict_avg8x8_smooth_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+#define vp9_eighttap_predict_avg8x8_smooth vp9_eighttap_predict_avg8x8_smooth_c
+
+void vp9_eighttap_predict_avg4x4_smooth_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+#define vp9_eighttap_predict_avg4x4_smooth vp9_eighttap_predict_avg4x4_smooth_c
+
+void vp9_eighttap_predict8x4_smooth_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+#define vp9_eighttap_predict8x4_smooth vp9_eighttap_predict8x4_smooth_c
+
+void vp9_eighttap_predict4x4_smooth_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+#define vp9_eighttap_predict4x4_smooth vp9_eighttap_predict4x4_smooth_c
+
+void vp9_sixtap_predict16x16_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_sixtap_predict16x16 vp9_sixtap_predict16x16_c
 
-void vp9_sixtap_predict8x8_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_sixtap_predict8x8_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_sixtap_predict8x8 vp9_sixtap_predict8x8_c
 
-void vp9_sixtap_predict_avg16x16_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_sixtap_predict_avg16x16_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_sixtap_predict_avg16x16 vp9_sixtap_predict_avg16x16_c
 
-void vp9_sixtap_predict_avg8x8_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_sixtap_predict_avg8x8_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_sixtap_predict_avg8x8 vp9_sixtap_predict_avg8x8_c
 
-void vp9_sixtap_predict8x4_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_sixtap_predict8x4_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_sixtap_predict8x4 vp9_sixtap_predict8x4_c
 
-void vp9_sixtap_predict_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-#define vp9_sixtap_predict vp9_sixtap_predict_c
-
-void vp9_sixtap_predict_avg_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-#define vp9_sixtap_predict_avg vp9_sixtap_predict_avg_c
-
-void vp9_bilinear_predict16x16_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-void vp9_bilinear_predict16x16_mmx(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-void vp9_bilinear_predict16x16_sse2(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-RTCD_EXTERN void (*vp9_bilinear_predict16x16)(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-
-void vp9_bilinear_predict8x8_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-void vp9_bilinear_predict8x8_mmx(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-void vp9_bilinear_predict8x8_sse2(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-RTCD_EXTERN void (*vp9_bilinear_predict8x8)(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-
-void vp9_bilinear_predict_avg16x16_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_sixtap_predict4x4_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+#define vp9_sixtap_predict4x4 vp9_sixtap_predict4x4_c
+
+void vp9_sixtap_predict_avg4x4_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+#define vp9_sixtap_predict_avg4x4 vp9_sixtap_predict_avg4x4_c
+
+void vp9_bilinear_predict16x16_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+void vp9_bilinear_predict16x16_sse2(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+RTCD_EXTERN void (*vp9_bilinear_predict16x16)(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+
+void vp9_bilinear_predict8x8_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+void vp9_bilinear_predict8x8_sse2(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+RTCD_EXTERN void (*vp9_bilinear_predict8x8)(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+
+void vp9_bilinear_predict_avg16x16_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_bilinear_predict_avg16x16 vp9_bilinear_predict_avg16x16_c
 
-void vp9_bilinear_predict_avg8x8_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_bilinear_predict_avg8x8_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_bilinear_predict_avg8x8 vp9_bilinear_predict_avg8x8_c
 
-void vp9_bilinear_predict8x4_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-void vp9_bilinear_predict8x4_mmx(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-RTCD_EXTERN void (*vp9_bilinear_predict8x4)(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-
-void vp9_bilinear_predict4x4_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-void vp9_bilinear_predict4x4_mmx(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-RTCD_EXTERN void (*vp9_bilinear_predict4x4)(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
-
-void vp9_bilinear_predict_avg4x4_c(unsigned char *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, unsigned char *dst_ptr, int  dst_pitch);
+void vp9_bilinear_predict8x4_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+#define vp9_bilinear_predict8x4 vp9_bilinear_predict8x4_c
+
+void vp9_bilinear_predict4x4_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
+#define vp9_bilinear_predict4x4 vp9_bilinear_predict4x4_c
+
+void vp9_bilinear_predict_avg4x4_c(uint8_t *src_ptr, int  src_pixels_per_line, int  xoffset, int  yoffset, uint8_t *dst_ptr, int  dst_pitch);
 #define vp9_bilinear_predict_avg4x4 vp9_bilinear_predict_avg4x4_c
 
-void vp9_short_idct4x4llm_1_c(short *input, short *output, int pitch);
+void vp9_short_idct4x4llm_1_c(int16_t *input, int16_t *output, int pitch);
 #define vp9_short_idct4x4llm_1 vp9_short_idct4x4llm_1_c
 
-void vp9_short_idct4x4llm_c(short *input, short *output, int pitch);
+void vp9_short_idct4x4llm_c(int16_t *input, int16_t *output, int pitch);
 #define vp9_short_idct4x4llm vp9_short_idct4x4llm_c
 
-void vp9_short_idct8x8_c(short *input, short *output, int pitch);
+void vp9_short_idct8x8_c(int16_t *input, int16_t *output, int pitch);
 #define vp9_short_idct8x8 vp9_short_idct8x8_c
 
-void vp9_short_idct10_8x8_c(short *input, short *output, int pitch);
+void vp9_short_idct10_8x8_c(int16_t *input, int16_t *output, int pitch);
 #define vp9_short_idct10_8x8 vp9_short_idct10_8x8_c
 
-void vp9_short_ihaar2x2_c(short *input, short *output, int pitch);
+void vp9_short_ihaar2x2_c(int16_t *input, int16_t *output, int pitch);
 #define vp9_short_ihaar2x2 vp9_short_ihaar2x2_c
 
-void vp9_short_idct16x16_c(short *input, short *output, int pitch);
+void vp9_short_idct16x16_c(int16_t *input, int16_t *output, int pitch);
 #define vp9_short_idct16x16 vp9_short_idct16x16_c
 
-void vp9_short_idct10_16x16_c(short *input, short *output, int pitch);
+void vp9_short_idct10_16x16_c(int16_t *input, int16_t *output, int pitch);
 #define vp9_short_idct10_16x16 vp9_short_idct10_16x16_c
 
-void vp9_ihtllm_c(const short *input, short *output, int pitch, int tx_type, int tx_dim, short eobs);
+void vp9_short_idct32x32_c(int16_t *input, int16_t *output, int pitch);
+#define vp9_short_idct32x32 vp9_short_idct32x32_c
+
+void vp9_ihtllm_c(const int16_t *input, int16_t *output, int pitch, int tx_type, int tx_dim, int16_t eobs);
 #define vp9_ihtllm vp9_ihtllm_c
 
-void vp9_short_inv_walsh4x4_1_c(short *in, short *out);
+void vp9_short_inv_walsh4x4_1_c(int16_t *in, int16_t *out);
 #define vp9_short_inv_walsh4x4_1 vp9_short_inv_walsh4x4_1_c
 
-void vp9_short_inv_walsh4x4_c(short *in, short *out);
+void vp9_short_inv_walsh4x4_c(int16_t *in, int16_t *out);
 #define vp9_short_inv_walsh4x4 vp9_short_inv_walsh4x4_c
 
-void vp9_dc_only_idct_add_8x8_c(short input_dc, unsigned char *pred_ptr, unsigned char *dst_ptr, int pitch, int stride);
+void vp9_dc_only_idct_add_8x8_c(int input_dc, uint8_t *pred_ptr, uint8_t *dst_ptr, int pitch, int stride);
 #define vp9_dc_only_idct_add_8x8 vp9_dc_only_idct_add_8x8_c
 
-void vp9_dc_only_idct_add_c(short input_dc, unsigned char *pred_ptr, unsigned char *dst_ptr, int pitch, int stride);
+void vp9_dc_only_idct_add_c(int input_dc, uint8_t *pred_ptr, uint8_t *dst_ptr, int pitch, int stride);
 #define vp9_dc_only_idct_add vp9_dc_only_idct_add_c
 
+unsigned int vp9_sad32x3_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int ref_stride, int max_sad);
+#define vp9_sad32x3 vp9_sad32x3_c
+
+unsigned int vp9_sad3x32_c(const uint8_t *src_ptr, int  src_stride, const uint8_t *ref_ptr, int ref_stride, int max_sad);
+#define vp9_sad3x32 vp9_sad3x32_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;
 
@@ skipping 15 matching lines @@
 
 
 
 
 
 
 
 
 
 
+
+
     vp9_copy_mem16x16 = vp9_copy_mem16x16_c;
     if (flags & HAS_MMX) vp9_copy_mem16x16 = vp9_copy_mem16x16_mmx;
     if (flags & HAS_SSE2) vp9_copy_mem16x16 = vp9_copy_mem16x16_sse2;
 
     vp9_copy_mem8x8 = vp9_copy_mem8x8_c;
     if (flags & HAS_MMX) vp9_copy_mem8x8 = vp9_copy_mem8x8_mmx;
 
     vp9_copy_mem8x4 = vp9_copy_mem8x4_c;
     if (flags & HAS_MMX) vp9_copy_mem8x4 = vp9_copy_mem8x4_mmx;
 
@@ skipping 21 matching lines @@
 
 
 
 
 
 
 
 
 
 
-
     vp9_loop_filter_mbv = vp9_loop_filter_mbv_c;
     if (flags & HAS_SSE2) vp9_loop_filter_mbv = vp9_loop_filter_mbv_sse2;
 
     vp9_loop_filter_bv = vp9_loop_filter_bv_c;
     if (flags & HAS_SSE2) vp9_loop_filter_bv = vp9_loop_filter_bv_sse2;
 
     vp9_loop_filter_bv8x8 = vp9_loop_filter_bv8x8_c;
     if (flags & HAS_SSE2) vp9_loop_filter_bv8x8 = vp9_loop_filter_bv8x8_sse2;
 
     vp9_loop_filter_mbh = vp9_loop_filter_mbh_c;
@@ skipping 14 matching lines @@
     if (flags & HAS_SSE2) vp9_loop_filter_simple_mbh = vp9_loop_filter_simple_horizontal_edge_sse2;
 
     vp9_loop_filter_simple_bv = vp9_loop_filter_bvs_c;
     if (flags & HAS_MMX) vp9_loop_filter_simple_bv = vp9_loop_filter_bvs_mmx;
     if (flags & HAS_SSE2) vp9_loop_filter_simple_bv = vp9_loop_filter_bvs_sse2;
 
     vp9_loop_filter_simple_bh = vp9_loop_filter_bhs_c;
     if (flags & HAS_MMX) vp9_loop_filter_simple_bh = vp9_loop_filter_bhs_mmx;
     if (flags & HAS_SSE2) vp9_loop_filter_simple_bh = vp9_loop_filter_bhs_sse2;
 
+    vp9_lpf_mbh_w = vp9_lpf_mbh_w_c;
+    if (flags & HAS_SSE2) vp9_lpf_mbh_w = vp9_lpf_mbh_w_sse2;
+
+    vp9_lpf_mbv_w = vp9_lpf_mbv_w_c;
+    if (flags & HAS_SSE2) vp9_lpf_mbv_w = vp9_lpf_mbv_w_sse2;
+
     vp9_mbpost_proc_down = vp9_mbpost_proc_down_c;
     if (flags & HAS_MMX) vp9_mbpost_proc_down = vp9_mbpost_proc_down_mmx;
     if (flags & HAS_SSE2) vp9_mbpost_proc_down = vp9_mbpost_proc_down_xmm;
 
     vp9_mbpost_proc_across_ip = vp9_mbpost_proc_across_ip_c;
     if (flags & HAS_SSE2) vp9_mbpost_proc_across_ip = vp9_mbpost_proc_across_ip_xmm;
 
     vp9_post_proc_down_and_across = vp9_post_proc_down_and_across_c;
     if (flags & HAS_MMX) vp9_post_proc_down_and_across = vp9_post_proc_down_and_across_mmx;
     if (flags & HAS_SSE2) vp9_post_proc_down_and_across = vp9_post_proc_down_and_across_xmm;
 
     vp9_plane_add_noise = vp9_plane_add_noise_c;
     if (flags & HAS_MMX) vp9_plane_add_noise = vp9_plane_add_noise_mmx;
     if (flags & HAS_SSE2) vp9_plane_add_noise = vp9_plane_add_noise_wmt;
 
 
 
 
     vp9_sad16x3 = vp9_sad16x3_c;
     if (flags & HAS_SSE2) vp9_sad16x3 = vp9_sad16x3_sse2;
 
     vp9_sad3x16 = vp9_sad3x16_c;
     if (flags & HAS_SSE2) vp9_sad3x16 = vp9_sad3x16_sse2;
 
+    vp9_sub_pixel_variance16x2 = vp9_sub_pixel_variance16x2_c;
+    if (flags & HAS_SSE2) vp9_sub_pixel_variance16x2 = vp9_sub_pixel_variance16x2_sse2;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
+
+
+
+
+
+
+
+
     vp9_bilinear_predict16x16 = vp9_bilinear_predict16x16_c;
-    if (flags & HAS_MMX) vp9_bilinear_predict16x16 = vp9_bilinear_predict16x16_mmx;
     if (flags & HAS_SSE2) vp9_bilinear_predict16x16 = vp9_bilinear_predict16x16_sse2;
 
     vp9_bilinear_predict8x8 = vp9_bilinear_predict8x8_c;
-    if (flags & HAS_MMX) vp9_bilinear_predict8x8 = vp9_bilinear_predict8x8_mmx;
     if (flags & HAS_SSE2) vp9_bilinear_predict8x8 = vp9_bilinear_predict8x8_sse2;
-
-
-
-    vp9_bilinear_predict8x4 = vp9_bilinear_predict8x4_c;
-    if (flags & HAS_MMX) vp9_bilinear_predict8x4 = vp9_bilinear_predict8x4_mmx;
-
-    vp9_bilinear_predict4x4 = vp9_bilinear_predict4x4_c;
-    if (flags & HAS_MMX) vp9_bilinear_predict4x4 = vp9_bilinear_predict4x4_mmx;
 }
 #endif
 #endif