1 #ifndef VP8_RTCD_H_
2 #define VP8_RTCD_H_
3
4 #ifdef RTCD_C
5 #define RTCD_EXTERN
6 #else
7 #define RTCD_EXTERN extern
8 #endif
9
10 /*
11 * VP8
12 */
13
14 struct blockd;
15 struct macroblockd;
16 struct loop_filter_info;
17
18 /* Encoder forward decls */
19 struct block;
20 struct macroblock;
21 struct variance_vtable;
22 union int_mv;
23 struct yv12_buffer_config;
24
25 #ifdef __cplusplus
26 extern "C" {
27 #endif
28
29 void vp8_bilinear_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
30 void vp8_bilinear_predict16x16_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
31 void vp8_bilinear_predict16x16_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
32 void vp8_bilinear_predict16x16_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
33 #define vp8_bilinear_predict16x16 vp8_bilinear_predict16x16_ssse3
34
35 void vp8_bilinear_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
36 void vp8_bilinear_predict4x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
37 #define vp8_bilinear_predict4x4 vp8_bilinear_predict4x4_mmx
38
39 void vp8_bilinear_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
40 void vp8_bilinear_predict8x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
41 #define vp8_bilinear_predict8x4 vp8_bilinear_predict8x4_mmx
42
43 void vp8_bilinear_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
44 void vp8_bilinear_predict8x8_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
45 void vp8_bilinear_predict8x8_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
46 void vp8_bilinear_predict8x8_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
47 #define vp8_bilinear_predict8x8 vp8_bilinear_predict8x8_ssse3
48
49 void vp8_blend_b_c(unsigned char *y, unsigned char *u, unsigned char *v, int y1, int u1, int v1, int alpha, int stride);
50 #define vp8_blend_b vp8_blend_b_c
51
52 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);
53 #define vp8_blend_mb_inner vp8_blend_mb_inner_c
54
55 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);
56 #define vp8_blend_mb_outer vp8_blend_mb_outer_c
57
58 int vp8_block_error_c(short *coeff, short *dqcoeff);
59 int vp8_block_error_mmx(short *coeff, short *dqcoeff);
60 int vp8_block_error_xmm(short *coeff, short *dqcoeff);
61 #define vp8_block_error vp8_block_error_xmm
62
63 void vp8_clear_system_state_c();
64 void vpx_reset_mmx_state();
65 #define vp8_clear_system_state vpx_reset_mmx_state
66
67 void vp8_copy32xn_c(const unsigned char *src_ptr, int source_stride, unsigned char *dst_ptr, int dst_stride, int n);
68 void vp8_copy32xn_sse2(const unsigned char *src_ptr, int source_stride, unsigned char *dst_ptr, int dst_stride, int n);
69 void vp8_copy32xn_sse3(const unsigned char *src_ptr, int source_stride, unsigned char *dst_ptr, int dst_stride, int n);
70 #define vp8_copy32xn vp8_copy32xn_sse3
71
72 void vp8_copy_mem16x16_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
73 void vp8_copy_mem16x16_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
74 void vp8_copy_mem16x16_sse2(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
75 #define vp8_copy_mem16x16 vp8_copy_mem16x16_sse2
76
77 void vp8_copy_mem8x4_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
78 void vp8_copy_mem8x4_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
79 #define vp8_copy_mem8x4 vp8_copy_mem8x4_mmx
80
81 void vp8_copy_mem8x8_c(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
82 void vp8_copy_mem8x8_mmx(unsigned char *src, int src_pitch, unsigned char *dst, int dst_pitch);
83 #define vp8_copy_mem8x8 vp8_copy_mem8x8_mmx
84
85 void vp8_dc_only_idct_add_c(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
86 void vp8_dc_only_idct_add_mmx(short input, unsigned char *pred, int pred_stride, unsigned char *dst, int dst_stride);
87 #define vp8_dc_only_idct_add vp8_dc_only_idct_add_mmx
88
89 int vp8_denoiser_filter_c(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
90 int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y, int mc_avg_y_stride, unsigned char *running_avg_y, int avg_y_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
91 #define vp8_denoiser_filter vp8_denoiser_filter_sse2
92
93 int vp8_denoiser_filter_uv_c(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
94 int vp8_denoiser_filter_uv_sse2(unsigned char *mc_running_avg, int mc_avg_stride, unsigned char *running_avg, int avg_stride, unsigned char *sig, int sig_stride, unsigned int motion_magnitude, int increase_denoising);
95 #define vp8_denoiser_filter_uv vp8_denoiser_filter_uv_sse2
96
97 void vp8_dequant_idct_add_c(short *input, short *dq, unsigned char *output, int stride);
98 void vp8_dequant_idct_add_mmx(short *input, short *dq, unsigned char *output, int stride);
99 #define vp8_dequant_idct_add vp8_dequant_idct_add_mmx
100
101 void vp8_dequant_idct_add_uv_block_c(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
102 void vp8_dequant_idct_add_uv_block_mmx(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
103 void vp8_dequant_idct_add_uv_block_sse2(short *q, short *dq, unsigned char *dst_u, unsigned char *dst_v, int stride, char *eobs);
104 #define vp8_dequant_idct_add_uv_block vp8_dequant_idct_add_uv_block_sse2
105
106 void vp8_dequant_idct_add_y_block_c(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
107 void vp8_dequant_idct_add_y_block_mmx(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
108 void vp8_dequant_idct_add_y_block_sse2(short *q, short *dq, unsigned char *dst, int stride, char *eobs);
109 #define vp8_dequant_idct_add_y_block vp8_dequant_idct_add_y_block_sse2
110
111 void vp8_dequantize_b_c(struct blockd*, short *dqc);
112 void vp8_dequantize_b_mmx(struct blockd*, short *dqc);
113 #define vp8_dequantize_b vp8_dequantize_b_mmx
114
115 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);
116 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);
117 #define vp8_diamond_search_sad vp8_diamond_search_sadx4
118
119 void vp8_fast_quantize_b_c(struct block *, struct blockd *);
120 void vp8_fast_quantize_b_sse2(struct block *, struct blockd *);
121 void vp8_fast_quantize_b_ssse3(struct block *, struct blockd *);
122 #define vp8_fast_quantize_b vp8_fast_quantize_b_ssse3
123
124 void vp8_filter_by_weight16x16_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
125 void vp8_filter_by_weight16x16_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
126 #define vp8_filter_by_weight16x16 vp8_filter_by_weight16x16_sse2
127
128 void vp8_filter_by_weight4x4_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
129 #define vp8_filter_by_weight4x4 vp8_filter_by_weight4x4_c
130
131 void vp8_filter_by_weight8x8_c(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
132 void vp8_filter_by_weight8x8_sse2(unsigned char *src, int src_stride, unsigned char *dst, int dst_stride, int src_weight);
133 #define vp8_filter_by_weight8x8 vp8_filter_by_weight8x8_sse2
134
135 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);
136 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);
137 #define vp8_full_search_sad vp8_full_search_sadx3
138
139 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);
140 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);
141 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);
142 #define vp8_loop_filter_bh vp8_loop_filter_bh_sse2
143
144 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);
145 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);
146 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);
147 #define vp8_loop_filter_bv vp8_loop_filter_bv_sse2
148
149 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);
150 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);
151 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);
152 #define vp8_loop_filter_mbh vp8_loop_filter_mbh_sse2
153
154 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);
155 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);
156 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);
157 #define vp8_loop_filter_mbv vp8_loop_filter_mbv_sse2
158
159 void vp8_loop_filter_bhs_c(unsigned char *y, int ystride, const unsigned char *blimit);
160 void vp8_loop_filter_bhs_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
161 void vp8_loop_filter_bhs_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
162 #define vp8_loop_filter_simple_bh vp8_loop_filter_bhs_sse2
163
164 void vp8_loop_filter_bvs_c(unsigned char *y, int ystride, const unsigned char *blimit);
165 void vp8_loop_filter_bvs_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
166 void vp8_loop_filter_bvs_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
167 #define vp8_loop_filter_simple_bv vp8_loop_filter_bvs_sse2
168
169 void vp8_loop_filter_simple_horizontal_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
170 void vp8_loop_filter_simple_horizontal_edge_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
171 void vp8_loop_filter_simple_horizontal_edge_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
172 #define vp8_loop_filter_simple_mbh vp8_loop_filter_simple_horizontal_edge_sse2
173
174 void vp8_loop_filter_simple_vertical_edge_c(unsigned char *y, int ystride, const unsigned char *blimit);
175 void vp8_loop_filter_simple_vertical_edge_mmx(unsigned char *y, int ystride, const unsigned char *blimit);
176 void vp8_loop_filter_simple_vertical_edge_sse2(unsigned char *y, int ystride, const unsigned char *blimit);
177 #define vp8_loop_filter_simple_mbv vp8_loop_filter_simple_vertical_edge_sse2
178
179 int vp8_mbblock_error_c(struct macroblock *mb, int dc);
180 int vp8_mbblock_error_mmx(struct macroblock *mb, int dc);
181 int vp8_mbblock_error_xmm(struct macroblock *mb, int dc);
182 #define vp8_mbblock_error vp8_mbblock_error_xmm
183
184 void vp8_mbpost_proc_across_ip_c(unsigned char *dst, int pitch, int rows, int cols,int flimit);
185 void vp8_mbpost_proc_across_ip_xmm(unsigned char *dst, int pitch, int rows, int cols,int flimit);
186 #define vp8_mbpost_proc_across_ip vp8_mbpost_proc_across_ip_xmm
187
188 void vp8_mbpost_proc_down_c(unsigned char *dst, int pitch, int rows, int cols,int flimit);
189 void vp8_mbpost_proc_down_mmx(unsigned char *dst, int pitch, int rows, int cols,int flimit);
190 void vp8_mbpost_proc_down_xmm(unsigned char *dst, int pitch, int rows, int cols,int flimit);
191 #define vp8_mbpost_proc_down vp8_mbpost_proc_down_xmm
192
193 int vp8_mbuverror_c(struct macroblock *mb);
194 int vp8_mbuverror_mmx(struct macroblock *mb);
195 int vp8_mbuverror_xmm(struct macroblock *mb);
196 #define vp8_mbuverror vp8_mbuverror_xmm
197
198 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);
199 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);
200 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);
201 #define vp8_plane_add_noise vp8_plane_add_noise_wmt
202
203 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);
204 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);
205 #define vp8_post_proc_down_and_across_mb_row vp8_post_proc_down_and_across_mb_row_sse2
206
207 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);
208 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);
209 #define vp8_refining_search_sad vp8_refining_search_sadx4
210
211 void vp8_regular_quantize_b_c(struct block *, struct blockd *);
212 void vp8_regular_quantize_b_sse2(struct block *, struct blockd *);
213 #define vp8_regular_quantize_b vp8_regular_quantize_b_sse2
214
215 void vp8_short_fdct4x4_c(short *input, short *output, int pitch);
216 void vp8_short_fdct4x4_mmx(short *input, short *output, int pitch);
217 void vp8_short_fdct4x4_sse2(short *input, short *output, int pitch);
218 #define vp8_short_fdct4x4 vp8_short_fdct4x4_sse2
219
220 void vp8_short_fdct8x4_c(short *input, short *output, int pitch);
221 void vp8_short_fdct8x4_mmx(short *input, short *output, int pitch);
222 void vp8_short_fdct8x4_sse2(short *input, short *output, int pitch);
223 #define vp8_short_fdct8x4 vp8_short_fdct8x4_sse2
224
225 void vp8_short_idct4x4llm_c(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
226 void vp8_short_idct4x4llm_mmx(short *input, unsigned char *pred, int pitch, unsigned char *dst, int dst_stride);
227 #define vp8_short_idct4x4llm vp8_short_idct4x4llm_mmx
228
229 void vp8_short_inv_walsh4x4_c(short *input, short *output);
230 void vp8_short_inv_walsh4x4_mmx(short *input, short *output);
231 void vp8_short_inv_walsh4x4_sse2(short *input, short *output);
232 #define vp8_short_inv_walsh4x4 vp8_short_inv_walsh4x4_sse2
233
234 void vp8_short_inv_walsh4x4_1_c(short *input, short *output);
235 #define vp8_short_inv_walsh4x4_1 vp8_short_inv_walsh4x4_1_c
236
237 void vp8_short_walsh4x4_c(short *input, short *output, int pitch);
238 void vp8_short_walsh4x4_sse2(short *input, short *output, int pitch);
239 #define vp8_short_walsh4x4 vp8_short_walsh4x4_sse2
240
241 void vp8_sixtap_predict16x16_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
242 void vp8_sixtap_predict16x16_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
243 void vp8_sixtap_predict16x16_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
244 void vp8_sixtap_predict16x16_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
245 #define vp8_sixtap_predict16x16 vp8_sixtap_predict16x16_ssse3
246
247 void vp8_sixtap_predict4x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
248 void vp8_sixtap_predict4x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
249 void vp8_sixtap_predict4x4_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
250 #define vp8_sixtap_predict4x4 vp8_sixtap_predict4x4_ssse3
251
252 void vp8_sixtap_predict8x4_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
253 void vp8_sixtap_predict8x4_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
254 void vp8_sixtap_predict8x4_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
255 void vp8_sixtap_predict8x4_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
256 #define vp8_sixtap_predict8x4 vp8_sixtap_predict8x4_ssse3
257
258 void vp8_sixtap_predict8x8_c(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
259 void vp8_sixtap_predict8x8_mmx(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
260 void vp8_sixtap_predict8x8_sse2(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
261 void vp8_sixtap_predict8x8_ssse3(unsigned char *src, int src_pitch, int xofst, int yofst, unsigned char *dst, int dst_pitch);
262 #define vp8_sixtap_predict8x8 vp8_sixtap_predict8x8_ssse3
263
264 void vp8_rtcd(void);
265
266 #ifdef RTCD_C
267 #include "vpx_ports/x86.h"
setup_rtcd_internal(void)268 static void setup_rtcd_internal(void)
269 {
270 int flags = x86_simd_caps();
271
272 (void)flags;
273
274 }
275 #endif
276
277 #ifdef __cplusplus
278 } // extern "C"
279 #endif
280
281 #endif
282