Home | History | Annotate | Download | only in x64 
      1 #ifndef VPX_RTCD_
      2 #define VPX_RTCD_
      3 
      4 #ifdef RTCD_C
      5 #define RTCD_EXTERN
      6 #else
      7 #define RTCD_EXTERN extern
      8 #endif
      9 
     10 #include "vp8/common/blockd.h"
     11 
     12 struct blockd;
     13 struct macroblockd;
     14 struct loop_filter_info;
     15 
     16 /* Encoder forward decls */
     17 struct block;
     18 struct macroblock;
     19 struct variance_vtable;
     20 union int_mv;
     21 struct yv12_buffer_config;
     22 
     23 void vp8_dequantize_b_c(struct blockd*, short *dqc);
     24 void vp8_dequantize_b_mmx(struct blockd*, short *dqc);
     25 #define vp8_dequantize_b vp8_dequantize_b_mmx
     26 
     27 void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
     28 void vp8_dequant_idct_add_mmx(short *input, short *dq, unsigned char *output, int stride);
     29 #define vp8_dequant_idct_add vp8_dequant_idct_add_mmx
     30 
     31 void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
     32 void vp8_dequant_idct_add_y_block_mmx(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
     33 void vp8_dequant_idct_add_y_block_sse2(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
     34 #define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_sse2
     35 
     36 void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
     37 void vp8_dequant_idct_add_uv_block_mmx(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
     38 void vp8_dequant_idct_add_uv_block_sse2(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
     39 #define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_sse2
     40 
     41 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);
     42 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);
     43 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);
     44 #define vp8_loop_filter_mbv vp8_loop_filter_mbv_sse2
     45 
     46 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);
     47 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);
     48 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);
     49 #define vp8_loop_filter_bv vp8_loop_filter_bv_sse2
     50 
     51 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);
     52 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);
     53 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);
     54 #define vp8_loop_filter_mbh vp8_loop_filter_mbh_sse2
     55 
     56 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);
     57 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);
     58 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);
     59 #define vp8_loop_filter_bh vp8_loop_filter_bh_sse2
     60 
     61 void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
     62 void vp8_loop_filter_simple_vertical_edge_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
     63 void vp8_loop_filter_simple_vertical_edge_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
     64 #define vp8_loop_filter_simple_mbv vp8_loop_filter_simple_vertical_edge_sse2
     65 
     66 void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
     67 void vp8_loop_filter_simple_horizontal_edge_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
     68 void vp8_loop_filter_simple_horizontal_edge_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
     69 #define vp8_loop_filter_simple_mbh vp8_loop_filter_simple_horizontal_edge_sse2
     70 
     71 void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
     72 void vp8_loop_filter_bvs_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
     73 void vp8_loop_filter_bvs_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
     74 #define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_sse2
     75 
     76 void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
     77 void vp8_loop_filter_bhs_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
     78 void vp8_loop_filter_bhs_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
     79 #define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_sse2
     80 
     81 void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
     82 void vp8_short_idct4x4llm_mmx(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
     83 #define vp8_short_idct4x4llm vp8_short_idct4x4llm_mmx
     84 
     85 void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
     86 #define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
     87 
     88 void vp8_short_inv_walsh4x4_c(short *input, short *output);
     89 void vp8_short_inv_walsh4x4_mmx(short *input, short *output);
     90 void vp8_short_inv_walsh4x4_sse2(short *input, short *output);
     91 #define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_sse2
     92 
     93 void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
     94 void vp8_dc_only_idct_add_mmx(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
     95 #define vp8_dc_only_idct_add vp8_dc_only_idct_add_mmx
     96 
     97 void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
     98 void vp8_copy_mem16x16_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
     99 void vp8_copy_mem16x16_sse2(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
    100 #define vp8_copy_mem16x16 vp8_copy_mem16x16_sse2
    101 
    102 void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
    103 void vp8_copy_mem8x8_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
    104 #define vp8_copy_mem8x8 vp8_copy_mem8x8_mmx
    105 
    106 void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
    107 void vp8_copy_mem8x4_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
    108 #define vp8_copy_mem8x4 vp8_copy_mem8x4_mmx
    109 
    110 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);
    111 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);
    112 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);
    113 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);
    114 
    115 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);
    116 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);
    117 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);
    118 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);
    119 
    120 void vp8_intra4x4_predict_c(unsigned char *Above, unsigned char *yleft, int left_stride, B_PREDICTION_MODE b_mode, unsigned char *dst, int dst_stride, unsigned char top_left);
    121 #define vp8_intra4x4_predict vp8_intra4x4_predict_c
    122 
    123 void vp8_mbpost_proc_down_c(unsigned char *dst, int pitch, int rows, int cols,int flimit);
    124 void vp8_mbpost_proc_down_mmx(unsigned char *dst, int pitch, int rows, int cols,int flimit);
    125 void vp8_mbpost_proc_down_xmm(unsigned char *dst, int pitch, int rows, int cols,int flimit);
    126 #define vp8_mbpost_proc_down vp8_mbpost_proc_down_xmm
    127 
    128 void vp8_mbpost_proc_across_ip_c(unsigned char *dst, int pitch, int rows, int cols,int flimit);
    129 void vp8_mbpost_proc_across_ip_xmm(unsigned char *dst, int pitch, int rows, int cols,int flimit);
    130 #define vp8_mbpost_proc_across_ip vp8_mbpost_proc_across_ip_xmm
    131 
    132 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);
    133 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);
    134 #define vp8_post_proc_down_and_across_mb_row vp8_post_proc_down_and_across_mb_row_sse2
    135 
    136 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);
    137 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);
    138 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);
    139 #define vp8_plane_add_noise vp8_plane_add_noise_wmt
    140 
    141 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);
    142 #define vp8_blend_mb_inner vp8_blend_mb_inner_c
    143 
    144 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);
    145 #define vp8_blend_mb_outer vp8_blend_mb_outer_c
    146 
    147 void vp8_blend_b_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
    148 #define vp8_blend_b vp8_blend_b_c
    149 
    150 void vp8_filter_by_weight16x16_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
    151 void vp8_filter_by_weight16x16_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
    152 #define vp8_filter_by_weight16x16 vp8_filter_by_weight16x16_sse2
    153 
    154 void vp8_filter_by_weight8x8_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
    155 void vp8_filter_by_weight8x8_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
    156 #define vp8_filter_by_weight8x8 vp8_filter_by_weight8x8_sse2
    157 
    158 void vp8_filter_by_weight4x4_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
    159 #define vp8_filter_by_weight4x4 vp8_filter_by_weight4x4_c
    160 
    161 void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    162 void vp8_sixtap_predict16x16_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    163 void vp8_sixtap_predict16x16_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    164 void vp8_sixtap_predict16x16_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    165 RTCD_EXTERN void (*vp8_sixtap_predict16x16)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    166 
    167 void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    168 void vp8_sixtap_predict8x8_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    169 void vp8_sixtap_predict8x8_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    170 void vp8_sixtap_predict8x8_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    171 RTCD_EXTERN void (*vp8_sixtap_predict8x8)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    172 
    173 void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    174 void vp8_sixtap_predict8x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    175 void vp8_sixtap_predict8x4_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    176 void vp8_sixtap_predict8x4_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    177 RTCD_EXTERN void (*vp8_sixtap_predict8x4)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    178 
    179 void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    180 void vp8_sixtap_predict4x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    181 void vp8_sixtap_predict4x4_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    182 RTCD_EXTERN void (*vp8_sixtap_predict4x4)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    183 
    184 void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    185 void vp8_bilinear_predict16x16_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    186 void vp8_bilinear_predict16x16_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    187 void vp8_bilinear_predict16x16_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    188 RTCD_EXTERN void (*vp8_bilinear_predict16x16)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    189 
    190 void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    191 void vp8_bilinear_predict8x8_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    192 void vp8_bilinear_predict8x8_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    193 void vp8_bilinear_predict8x8_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    194 RTCD_EXTERN void (*vp8_bilinear_predict8x8)(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    195 
    196 void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    197 void vp8_bilinear_predict8x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    198 #define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_mmx
    199 
    200 void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    201 void vp8_bilinear_predict4x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
    202 #define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_mmx
    203 
    204 unsigned int vp8_variance4x4_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    205 unsigned int vp8_variance4x4_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    206 unsigned int vp8_variance4x4_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    207 #define vp8_variance4x4 vp8_variance4x4_wmt
    208 
    209 unsigned int vp8_variance8x8_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    210 unsigned int vp8_variance8x8_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    211 unsigned int vp8_variance8x8_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    212 #define vp8_variance8x8 vp8_variance8x8_wmt
    213 
    214 unsigned int vp8_variance8x16_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    215 unsigned int vp8_variance8x16_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    216 unsigned int vp8_variance8x16_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    217 #define vp8_variance8x16 vp8_variance8x16_wmt
    218 
    219 unsigned int vp8_variance16x8_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    220 unsigned int vp8_variance16x8_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    221 unsigned int vp8_variance16x8_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    222 #define vp8_variance16x8 vp8_variance16x8_wmt
    223 
    224 unsigned int vp8_variance16x16_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    225 unsigned int vp8_variance16x16_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    226 unsigned int vp8_variance16x16_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    227 #define vp8_variance16x16 vp8_variance16x16_wmt
    228 
    229 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);
    230 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);
    231 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);
    232 #define vp8_sub_pixel_variance4x4 vp8_sub_pixel_variance4x4_wmt
    233 
    234 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);
    235 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);
    236 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);
    237 #define vp8_sub_pixel_variance8x8 vp8_sub_pixel_variance8x8_wmt
    238 
    239 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);
    240 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);
    241 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);
    242 #define vp8_sub_pixel_variance8x16 vp8_sub_pixel_variance8x16_wmt
    243 
    244 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);
    245 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);
    246 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);
    247 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);
    248 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);
    249 
    250 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);
    251 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);
    252 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);
    253 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);
    254 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);
    255 
    256 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);
    257 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);
    258 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);
    259 #define vp8_variance_halfpixvar16x16_h vp8_variance_halfpixvar16x16_h_wmt
    260 
    261 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);
    262 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);
    263 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);
    264 #define vp8_variance_halfpixvar16x16_v vp8_variance_halfpixvar16x16_v_wmt
    265 
    266 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);
    267 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);
    268 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);
    269 #define vp8_variance_halfpixvar16x16_hv vp8_variance_halfpixvar16x16_hv_wmt
    270 
    271 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);
    272 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);
    273 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);
    274 #define vp8_sad4x4 vp8_sad4x4_wmt
    275 
    276 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);
    277 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);
    278 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);
    279 #define vp8_sad8x8 vp8_sad8x8_wmt
    280 
    281 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);
    282 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);
    283 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);
    284 #define vp8_sad8x16 vp8_sad8x16_wmt
    285 
    286 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);
    287 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);
    288 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);
    289 #define vp8_sad16x8 vp8_sad16x8_wmt
    290 
    291 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);
    292 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);
    293 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);
    294 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);
    295 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);
    296 
    297 void vp8_sad4x4x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
    298 void vp8_sad4x4x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
    299 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);
    300 
    301 void vp8_sad8x8x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
    302 void vp8_sad8x8x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
    303 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);
    304 
    305 void vp8_sad8x16x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
    306 void vp8_sad8x16x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
    307 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);
    308 
    309 void vp8_sad16x8x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
    310 void vp8_sad16x8x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
    311 void vp8_sad16x8x3_ssse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
    312 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);
    313 
    314 void vp8_sad16x16x3_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
    315 void vp8_sad16x16x3_sse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
    316 void vp8_sad16x16x3_ssse3(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sad_array);
    317 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);
    318 
    319 void vp8_sad4x4x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
    320 void vp8_sad4x4x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
    321 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);
    322 
    323 void vp8_sad8x8x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
    324 void vp8_sad8x8x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
    325 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);
    326 
    327 void vp8_sad8x16x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
    328 void vp8_sad8x16x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
    329 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);
    330 
    331 void vp8_sad16x8x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
    332 void vp8_sad16x8x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
    333 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);
    334 
    335 void vp8_sad16x16x8_c(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
    336 void vp8_sad16x16x8_sse4(const unsigned char *src_ptr, int src_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned short *sad_array);
    337 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);
    338 
    339 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);
    340 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);
    341 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);
    342 
    343 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);
    344 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);
    345 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);
    346 
    347 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);
    348 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);
    349 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);
    350 
    351 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);
    352 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);
    353 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);
    354 
    355 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);
    356 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);
    357 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);
    358 
    359 unsigned int vp8_get_mb_ss_c(const short *);
    360 unsigned int vp8_get_mb_ss_mmx(const short *);
    361 unsigned int vp8_get_mb_ss_sse2(const short *);
    362 #define vp8_get_mb_ss vp8_get_mb_ss_sse2
    363 
    364 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);
    365 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);
    366 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);
    367 #define vp8_sub_pixel_mse16x16 vp8_sub_pixel_mse16x16_wmt
    368 
    369 unsigned int vp8_mse16x16_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    370 unsigned int vp8_mse16x16_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    371 unsigned int vp8_mse16x16_wmt(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride, unsigned int *sse);
    372 #define vp8_mse16x16 vp8_mse16x16_wmt
    373 
    374 unsigned int vp8_get4x4sse_cs_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
    375 unsigned int vp8_get4x4sse_cs_mmx(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int  ref_stride);
    376 #define vp8_get4x4sse_cs vp8_get4x4sse_cs_mmx
    377 
    378 void vp8_copy32xn_c(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, int n);
    379 void vp8_copy32xn_sse2(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, int n);
    380 void vp8_copy32xn_sse3(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, int n);
    381 RTCD_EXTERN void (*vp8_copy32xn)(const unsigned char *src_ptr, int source_stride, const unsigned char *ref_ptr, int ref_stride, int n);
    382 
    383 void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
    384 void vp8_short_fdct4x4_mmx(short *input, short *output, int pitch);
    385 void vp8_short_fdct4x4_sse2(short *input, short *output, int pitch);
    386 #define vp8_short_fdct4x4 vp8_short_fdct4x4_sse2
    387 
    388 void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
    389 void vp8_short_fdct8x4_mmx(short *input, short *output, int pitch);
    390 void vp8_short_fdct8x4_sse2(short *input, short *output, int pitch);
    391 #define vp8_short_fdct8x4 vp8_short_fdct8x4_sse2
    392 
    393 void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
    394 void vp8_short_walsh4x4_sse2(short *input, short *output, int pitch);
    395 #define vp8_short_walsh4x4 vp8_short_walsh4x4_sse2
    396 
    397 void vp8_regular_quantize_b_c(struct block *, struct blockd *);
    398 void vp8_regular_quantize_b_sse2(struct block *, struct blockd *);
    399 void vp8_regular_quantize_b_sse4(struct block *, struct blockd *);
    400 RTCD_EXTERN void (*vp8_regular_quantize_b)(struct block *, struct blockd *);
    401 
    402 void vp8_fast_quantize_b_c(struct block *, struct blockd *);
    403 void vp8_fast_quantize_b_sse2(struct block *, struct blockd *);
    404 void vp8_fast_quantize_b_ssse3(struct block *, struct blockd *);
    405 RTCD_EXTERN void (*vp8_fast_quantize_b)(struct block *, struct blockd *);
    406 
    407 void vp8_regular_quantize_b_pair_c(struct block *b1, struct block *b2, struct blockd *d1, struct blockd *d2);
    408 #define vp8_regular_quantize_b_pair vp8_regular_quantize_b_pair_c
    409 
    410 void vp8_fast_quantize_b_pair_c(struct block *b1, struct block *b2, struct blockd *d1, struct blockd *d2);
    411 #define vp8_fast_quantize_b_pair vp8_fast_quantize_b_pair_c
    412 
    413 void vp8_quantize_mb_c(struct macroblock *);
    414 #define vp8_quantize_mb vp8_quantize_mb_c
    415 
    416 void vp8_quantize_mby_c(struct macroblock *);
    417 #define vp8_quantize_mby vp8_quantize_mby_c
    418 
    419 void vp8_quantize_mbuv_c(struct macroblock *);
    420 #define vp8_quantize_mbuv vp8_quantize_mbuv_c
    421 
    422 int vp8_block_error_c(short *coeff, short *dqcoeff);
    423 int vp8_block_error_mmx(short *coeff, short *dqcoeff);
    424 int vp8_block_error_xmm(short *coeff, short *dqcoeff);
    425 #define vp8_block_error vp8_block_error_xmm
    426 
    427 int vp8_mbblock_error_c(struct macroblock *mb, int dc);
    428 int vp8_mbblock_error_mmx(struct macroblock *mb, int dc);
    429 int vp8_mbblock_error_xmm(struct macroblock *mb, int dc);
    430 #define vp8_mbblock_error vp8_mbblock_error_xmm
    431 
    432 int vp8_mbuverror_c(struct macroblock *mb);
    433 int vp8_mbuverror_mmx(struct macroblock *mb);
    434 int vp8_mbuverror_xmm(struct macroblock *mb);
    435 #define vp8_mbuverror vp8_mbuverror_xmm
    436 
    437 void vp8_subtract_b_c(struct block *be, struct blockd *bd, int pitch);
    438 void vp8_subtract_b_mmx(struct block *be, struct blockd *bd, int pitch);
    439 void vp8_subtract_b_sse2(struct block *be, struct blockd *bd, int pitch);
    440 #define vp8_subtract_b vp8_subtract_b_sse2
    441 
    442 void vp8_subtract_mby_c(short *diff, unsigned char *src, int src_stride, unsigned char *pred, int pred_stride);
    443 void vp8_subtract_mby_mmx(short *diff, unsigned char *src, int src_stride, unsigned char *pred, int pred_stride);
    444 void vp8_subtract_mby_sse2(short *diff, unsigned char *src, int src_stride, unsigned char *pred, int pred_stride);
    445 #define vp8_subtract_mby vp8_subtract_mby_sse2
    446 
    447 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);
    448 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);
    449 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);
    450 #define vp8_subtract_mbuv vp8_subtract_mbuv_sse2
    451 
    452 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);
    453 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);
    454 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);
    455 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);
    456 
    457 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);
    458 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);
    459 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);
    460 
    461 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);
    462 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);
    463 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);
    464 
    465 void vp8_yv12_copy_partial_frame_c(struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
    466 #define vp8_yv12_copy_partial_frame vp8_yv12_copy_partial_frame_c
    467 
    468 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);
    469 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);
    470 #define vp8_denoiser_filter vp8_denoiser_filter_sse2
    471 
    472 void vp8_horizontal_line_4_5_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
    473 #define vp8_horizontal_line_4_5_scale vp8_horizontal_line_4_5_scale_c
    474 
    475 void vp8_vertical_band_4_5_scale_c(unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
    476 #define vp8_vertical_band_4_5_scale vp8_vertical_band_4_5_scale_c
    477 
    478 void vp8_last_vertical_band_4_5_scale_c(unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
    479 #define vp8_last_vertical_band_4_5_scale vp8_last_vertical_band_4_5_scale_c
    480 
    481 void vp8_horizontal_line_2_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
    482 #define vp8_horizontal_line_2_3_scale vp8_horizontal_line_2_3_scale_c
    483 
    484 void vp8_vertical_band_2_3_scale_c(unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
    485 #define vp8_vertical_band_2_3_scale vp8_vertical_band_2_3_scale_c
    486 
    487 void vp8_last_vertical_band_2_3_scale_c(unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
    488 #define vp8_last_vertical_band_2_3_scale vp8_last_vertical_band_2_3_scale_c
    489 
    490 void vp8_horizontal_line_3_5_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
    491 #define vp8_horizontal_line_3_5_scale vp8_horizontal_line_3_5_scale_c
    492 
    493 void vp8_vertical_band_3_5_scale_c(unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
    494 #define vp8_vertical_band_3_5_scale vp8_vertical_band_3_5_scale_c
    495 
    496 void vp8_last_vertical_band_3_5_scale_c(unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
    497 #define vp8_last_vertical_band_3_5_scale vp8_last_vertical_band_3_5_scale_c
    498 
    499 void vp8_horizontal_line_3_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
    500 #define vp8_horizontal_line_3_4_scale vp8_horizontal_line_3_4_scale_c
    501 
    502 void vp8_vertical_band_3_4_scale_c(unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
    503 #define vp8_vertical_band_3_4_scale vp8_vertical_band_3_4_scale_c
    504 
    505 void vp8_last_vertical_band_3_4_scale_c(unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
    506 #define vp8_last_vertical_band_3_4_scale vp8_last_vertical_band_3_4_scale_c
    507 
    508 void vp8_horizontal_line_1_2_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
    509 #define vp8_horizontal_line_1_2_scale vp8_horizontal_line_1_2_scale_c
    510 
    511 void vp8_vertical_band_1_2_scale_c(unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
    512 #define vp8_vertical_band_1_2_scale vp8_vertical_band_1_2_scale_c
    513 
    514 void vp8_last_vertical_band_1_2_scale_c(unsigned char *dest, unsigned int dest_pitch, unsigned int dest_width);
    515 #define vp8_last_vertical_band_1_2_scale vp8_last_vertical_band_1_2_scale_c
    516 
    517 void vp8_horizontal_line_5_4_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
    518 #define vp8_horizontal_line_5_4_scale vp8_horizontal_line_5_4_scale_c
    519 
    520 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);
    521 #define vp8_vertical_band_5_4_scale vp8_vertical_band_5_4_scale_c
    522 
    523 void vp8_horizontal_line_5_3_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
    524 #define vp8_horizontal_line_5_3_scale vp8_horizontal_line_5_3_scale_c
    525 
    526 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);
    527 #define vp8_vertical_band_5_3_scale vp8_vertical_band_5_3_scale_c
    528 
    529 void vp8_horizontal_line_2_1_scale_c(const unsigned char *source, unsigned int source_width, unsigned char *dest, unsigned int dest_width);
    530 #define vp8_horizontal_line_2_1_scale vp8_horizontal_line_2_1_scale_c
    531 
    532 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);
    533 #define vp8_vertical_band_2_1_scale vp8_vertical_band_2_1_scale_c
    534 
    535 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);
    536 #define vp8_vertical_band_2_1_scale_i vp8_vertical_band_2_1_scale_i_c
    537 
    538 void vp8_yv12_extend_frame_borders_c(struct yv12_buffer_config *ybf);
    539 #define vp8_yv12_extend_frame_borders vp8_yv12_extend_frame_borders_c
    540 
    541 void vp8_yv12_copy_frame_c(struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
    542 #define vp8_yv12_copy_frame vp8_yv12_copy_frame_c
    543 
    544 void vp8_yv12_copy_y_c(struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc);
    545 #define vp8_yv12_copy_y vp8_yv12_copy_y_c
    546 
    547 void vpx_rtcd(void);
    548 
    549 #ifdef RTCD_C
    550 #include "vpx_ports/x86.h"
    551 static void setup_rtcd_internal(void)
    552 {
    553     int flags = x86_simd_caps();
    554 
    555     (void)flags;
    556 
    557 
    558 
    559 
    560 
    561 
    562 
    563 
    564 
    565 
    566 
    567 
    568 
    569 
    570 
    571 
    572 
    573 
    574 
    575 
    576     vp8_build_intra_predictors_mby_s = vp8_build_intra_predictors_mby_s_sse2;
    577     if (flags & HAS_SSSE3) vp8_build_intra_predictors_mby_s = vp8_build_intra_predictors_mby_s_ssse3;
    578 
    579     vp8_build_intra_predictors_mbuv_s = vp8_build_intra_predictors_mbuv_s_sse2;
    580     if (flags & HAS_SSSE3) vp8_build_intra_predictors_mbuv_s = vp8_build_intra_predictors_mbuv_s_ssse3;
    581 
    582 
    583 
    584 
    585 
    586 
    587 
    588 
    589 
    590 
    591 
    592 
    593     vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_sse2;
    594     if (flags & HAS_SSSE3) vp8_sixtap_predict16x16 = vp8_sixtap_predict16x16_ssse3;
    595 
    596     vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_sse2;
    597     if (flags & HAS_SSSE3) vp8_sixtap_predict8x8 = vp8_sixtap_predict8x8_ssse3;
    598 
    599     vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_sse2;
    600     if (flags & HAS_SSSE3) vp8_sixtap_predict8x4 = vp8_sixtap_predict8x4_ssse3;
    601 
    602     vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_mmx;
    603     if (flags & HAS_SSSE3) vp8_sixtap_predict4x4 = vp8_sixtap_predict4x4_ssse3;
    604 
    605     vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_sse2;
    606     if (flags & HAS_SSSE3) vp8_bilinear_predict16x16 = vp8_bilinear_predict16x16_ssse3;
    607 
    608     vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_sse2;
    609     if (flags & HAS_SSSE3) vp8_bilinear_predict8x8 = vp8_bilinear_predict8x8_ssse3;
    610 
    611 
    612 
    613 
    614 
    615 
    616 
    617 
    618 
    619 
    620 
    621     vp8_sub_pixel_variance16x8 = vp8_sub_pixel_variance16x8_wmt;
    622     if (flags & HAS_SSSE3) vp8_sub_pixel_variance16x8 = vp8_sub_pixel_variance16x8_ssse3;
    623 
    624     vp8_sub_pixel_variance16x16 = vp8_sub_pixel_variance16x16_wmt;
    625     if (flags & HAS_SSSE3) vp8_sub_pixel_variance16x16 = vp8_sub_pixel_variance16x16_ssse3;
    626 
    627 
    628 
    629 
    630 
    631 
    632 
    633 
    634     vp8_sad16x16 = vp8_sad16x16_wmt;
    635     if (flags & HAS_SSE3) vp8_sad16x16 = vp8_sad16x16_sse3;
    636 
    637     vp8_sad4x4x3 = vp8_sad4x4x3_c;
    638     if (flags & HAS_SSE3) vp8_sad4x4x3 = vp8_sad4x4x3_sse3;
    639 
    640     vp8_sad8x8x3 = vp8_sad8x8x3_c;
    641     if (flags & HAS_SSE3) vp8_sad8x8x3 = vp8_sad8x8x3_sse3;
    642 
    643     vp8_sad8x16x3 = vp8_sad8x16x3_c;
    644     if (flags & HAS_SSE3) vp8_sad8x16x3 = vp8_sad8x16x3_sse3;
    645 
    646     vp8_sad16x8x3 = vp8_sad16x8x3_c;
    647     if (flags & HAS_SSE3) vp8_sad16x8x3 = vp8_sad16x8x3_sse3;
    648     if (flags & HAS_SSSE3) vp8_sad16x8x3 = vp8_sad16x8x3_ssse3;
    649 
    650     vp8_sad16x16x3 = vp8_sad16x16x3_c;
    651     if (flags & HAS_SSE3) vp8_sad16x16x3 = vp8_sad16x16x3_sse3;
    652     if (flags & HAS_SSSE3) vp8_sad16x16x3 = vp8_sad16x16x3_ssse3;
    653 
    654     vp8_sad4x4x8 = vp8_sad4x4x8_c;
    655     if (flags & HAS_SSE4_1) vp8_sad4x4x8 = vp8_sad4x4x8_sse4;
    656 
    657     vp8_sad8x8x8 = vp8_sad8x8x8_c;
    658     if (flags & HAS_SSE4_1) vp8_sad8x8x8 = vp8_sad8x8x8_sse4;
    659 
    660     vp8_sad8x16x8 = vp8_sad8x16x8_c;
    661     if (flags & HAS_SSE4_1) vp8_sad8x16x8 = vp8_sad8x16x8_sse4;
    662 
    663     vp8_sad16x8x8 = vp8_sad16x8x8_c;
    664     if (flags & HAS_SSE4_1) vp8_sad16x8x8 = vp8_sad16x8x8_sse4;
    665 
    666     vp8_sad16x16x8 = vp8_sad16x16x8_c;
    667     if (flags & HAS_SSE4_1) vp8_sad16x16x8 = vp8_sad16x16x8_sse4;
    668 
    669     vp8_sad4x4x4d = vp8_sad4x4x4d_c;
    670     if (flags & HAS_SSE3) vp8_sad4x4x4d = vp8_sad4x4x4d_sse3;
    671 
    672     vp8_sad8x8x4d = vp8_sad8x8x4d_c;
    673     if (flags & HAS_SSE3) vp8_sad8x8x4d = vp8_sad8x8x4d_sse3;
    674 
    675     vp8_sad8x16x4d = vp8_sad8x16x4d_c;
    676     if (flags & HAS_SSE3) vp8_sad8x16x4d = vp8_sad8x16x4d_sse3;
    677 
    678     vp8_sad16x8x4d = vp8_sad16x8x4d_c;
    679     if (flags & HAS_SSE3) vp8_sad16x8x4d = vp8_sad16x8x4d_sse3;
    680 
    681     vp8_sad16x16x4d = vp8_sad16x16x4d_c;
    682     if (flags & HAS_SSE3) vp8_sad16x16x4d = vp8_sad16x16x4d_sse3;
    683 
    684 
    685 
    686 
    687 
    688     vp8_copy32xn = vp8_copy32xn_sse2;
    689     if (flags & HAS_SSE3) vp8_copy32xn = vp8_copy32xn_sse3;
    690 
    691 
    692 
    693 
    694     vp8_regular_quantize_b = vp8_regular_quantize_b_sse2;
    695     if (flags & HAS_SSE4_1) vp8_regular_quantize_b = vp8_regular_quantize_b_sse4;
    696 
    697     vp8_fast_quantize_b = vp8_fast_quantize_b_sse2;
    698     if (flags & HAS_SSSE3) vp8_fast_quantize_b = vp8_fast_quantize_b_ssse3;
    699 
    700 
    701 
    702 
    703 
    704 
    705 
    706 
    707 
    708 
    709 
    710 
    711     vp8_full_search_sad = vp8_full_search_sad_c;
    712     if (flags & HAS_SSE3) vp8_full_search_sad = vp8_full_search_sadx3;
    713     if (flags & HAS_SSE4_1) vp8_full_search_sad = vp8_full_search_sadx8;
    714 
    715     vp8_refining_search_sad = vp8_refining_search_sad_c;
    716     if (flags & HAS_SSE3) vp8_refining_search_sad = vp8_refining_search_sadx4;
    717 
    718     vp8_diamond_search_sad = vp8_diamond_search_sad_c;
    719     if (flags & HAS_SSE3) vp8_diamond_search_sad = vp8_diamond_search_sadx4;
    720 }
    721 #endif
    722 #endif
    723